Datasets:

---

ftopal

/

huggingface-models

like 0

[ "transformers", "safetensors", "mistral", "text-generation", "mergekit", "merge", "arxiv:2403.19522", "base_model:Undi95/Toppy-M-7B", "base_model:SanjiWatsuki/Kunoichi-DPO-v2-7B", "base_model:Epiculous/Fett-uccine-7B", "base_model:NeverSleep/Noromaid-7B-0.4-DPO", "license:apache-2.0", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "region:us" ]

[ "2403.19522" ]

[]

[ "# merge\n\nThis is a merge of pre-trained language models created using mergekit.", "## Merge Details", "### Merge Method\n\nThis model was merged using the Model Stock merge method using SanjiWatsuki/Kunoichi-DPO-v2-7B as a base.", "### Models Merged\n\nThe following models were included in the merge:\n* Undi95/Toppy-M-7B\n* Epiculous/Fett-uccine-7B\n* NeverSleep/Noromaid-7B-0.4-DPO", "### Configuration\n\nThe following YAML configuration was used to produce this model:" ]

[ "TAGS\n#transformers #safetensors #mistral #text-generation #mergekit #merge #arxiv-2403.19522 #base_model-Undi95/Toppy-M-7B #base_model-SanjiWatsuki/Kunoichi-DPO-v2-7B #base_model-Epiculous/Fett-uccine-7B #base_model-NeverSleep/Noromaid-7B-0.4-DPO #license-apache-2.0 #autotrain_compatible #endpoints_compatible #text-generation-inference #region-us \n", "# merge\n\nThis is a merge of pre-trained language models created using mergekit.", "## Merge Details", "### Merge Method\n\nThis model was merged using the Model Stock merge method using SanjiWatsuki/Kunoichi-DPO-v2-7B as a base.", "### Models Merged\n\nThe following models were included in the merge:\n* Undi95/Toppy-M-7B\n* Epiculous/Fett-uccine-7B\n* NeverSleep/Noromaid-7B-0.4-DPO", "### Configuration\n\nThe following YAML configuration was used to produce this model:" ]

[ "transformers", "safetensors", "llama", "text-generation", "conversational", "arxiv:1910.09700", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "region:us" ]

[ "1910.09700" ]

[]

[ "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a transformers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "TAGS\n#transformers #safetensors #llama #text-generation #conversational #arxiv-1910.09700 #autotrain_compatible #endpoints_compatible #text-generation-inference #region-us \n", "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a transformers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "diffusers", "safetensors", "arxiv:1910.09700", "endpoints_compatible", "diffusers:StableDiffusionPipeline", "region:us" ]

[ "1910.09700" ]

[]

[ "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a diffusers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "TAGS\n#diffusers #safetensors #arxiv-1910.09700 #endpoints_compatible #diffusers-StableDiffusionPipeline #region-us \n", "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a diffusers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2204.06745", "arxiv:2101.00027", "arxiv:2201.07311", "arxiv:2104.09864", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[ "2204.06745", "2101.00027", "2201.07311", "2104.09864" ]

[]

[ "### Model details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: GPT-NeoX-20B: An Open-Source Autoregressive Language\nModel. For details about the training dataset,\nsee the Pile paper, and its data\nsheet.\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing GPT-NeoX-20B documentation before asking about the model\non Discord. For general correspondence: contact@eleuther.\nai.", "### Uses and limitations", "#### Intended use\n\n\nGPT-NeoX-20B was developed primarily for research purposes. It learns an inner\nrepresentation of the English language that can be used to extract features\nuseful for downstream tasks.\n\n\nIn addition to scientific uses, you may also further fine-tune and adapt\nGPT-NeoX-20B for deployment, as long as your use is in accordance with the\nApache 2.0 license. This model works with the Transformers\nLibrary. If you decide to use\npre-trained GPT-NeoX-20B as a basis for your fine-tuned model, please note that\nyou need to conduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nGPT-NeoX-20B is not intended for deployment as-is. It is not a product\nand cannot be used for human-facing interactions without supervision.\n\n\nGPT-NeoX-20B has not been fine-tuned for downstream tasks for which language\nmodels are commonly deployed, such as writing genre prose, or commercial\nchatbots. This means GPT-NeoX-20B will likely not respond to a given prompt\nthe way products such as ChatGPT do. This is because, unlike GPT-NeoX-20B,\nChatGPT was fine-tuned using methods such as Reinforcement Learning from Human\nFeedback (RLHF) to better “understand” human instructions and dialogue.\n\n\nThis model is English-language only, and thus cannot be used for translation\nor generating text in other languages.", "#### Limitations and biases\n\n\nThe core functionality of GPT-NeoX-20B is to take a string of text and predict\nthe next token. Remember that the statistically most likely next token need\nnot result in the most “accurate” text. Never rely on GPT-NeoX-20B to produce\nfactually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nGPT-NeoX-20B may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nWe recommend curating the outputs of this model before presenting it to a human\nreader. Please inform your audience that you are using artificially generated\ntext.", "#### How to use\n\n\nIf you simply want to try out some prompts, check out this\nplayground.\n\n\nGPT-NeoX-20B can be loaded using the 'AutoModelForCausalLM' functionality:", "### Training", "#### Training dataset\n\n\nThe Pile is a 825GiB general-purpose dataset in English. It was created by\nEleutherAI specifically for training large language models. It contains texts\nfrom 22 diverse sources, roughly broken down into five categories: academic\nwriting (e.g. arXiv), internet (e.g. CommonCrawl), prose (e.g. Project\nGutenberg), dialogue (e.g. YouTube subtitles), and miscellaneous (e.g. GitHub,\nEnron Emails). See the Pile paper for\na breakdown of all data sources, methodology, and a discussion of ethical\nimplications. Consult the datasheet for\nmore detailed documentation about the Pile and its component datasets. The\nPile can be downloaded from the official website,\nor from a community mirror.\n\n\nThe Pile was not deduplicated before being used to train GPT-NeoX-20B.", "#### Training procedure\n\n\nGPT-NeoX-20B was trained with a batch size of approximately 3.15M tokens\n(1538 sequences of 2048 tokens each), for a total of 150,000 steps. Tensor\nparallelism and pipeline parallelism were used to distribute the model across\nGPUs. Additional details about the training procedure are in Section 3 of\nthe accompanying paper.", "### Evaluations\n\n\n\n\nZero-shot performance on selected natural language tasks.\n\nThis is a heavily abridged version of the evaluation results. Appendix D of the\nGPT-NeoX-20B paper compares more model\nsizes, and contains additional evaluations, including on: zero and five-shot\nnatural language tasks, zero and five-shot Basic Arithmetic and MATH,\nand zero-shot Hendrycks tasks.", "### BibTeX\n\n\nTo cite the GPT-NeoX-20B paper:\n\n\nOpen LLM Leaderboard Evaluation Results\n=======================================\n\n\nDetailed results can be found here" ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2204.06745 #arxiv-2101.00027 #arxiv-2201.07311 #arxiv-2104.09864 #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n", "### Model details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: GPT-NeoX-20B: An Open-Source Autoregressive Language\nModel. For details about the training dataset,\nsee the Pile paper, and its data\nsheet.\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing GPT-NeoX-20B documentation before asking about the model\non Discord. For general correspondence: contact@eleuther.\nai.", "### Uses and limitations", "#### Intended use\n\n\nGPT-NeoX-20B was developed primarily for research purposes. It learns an inner\nrepresentation of the English language that can be used to extract features\nuseful for downstream tasks.\n\n\nIn addition to scientific uses, you may also further fine-tune and adapt\nGPT-NeoX-20B for deployment, as long as your use is in accordance with the\nApache 2.0 license. This model works with the Transformers\nLibrary. If you decide to use\npre-trained GPT-NeoX-20B as a basis for your fine-tuned model, please note that\nyou need to conduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nGPT-NeoX-20B is not intended for deployment as-is. It is not a product\nand cannot be used for human-facing interactions without supervision.\n\n\nGPT-NeoX-20B has not been fine-tuned for downstream tasks for which language\nmodels are commonly deployed, such as writing genre prose, or commercial\nchatbots. This means GPT-NeoX-20B will likely not respond to a given prompt\nthe way products such as ChatGPT do. This is because, unlike GPT-NeoX-20B,\nChatGPT was fine-tuned using methods such as Reinforcement Learning from Human\nFeedback (RLHF) to better “understand” human instructions and dialogue.\n\n\nThis model is English-language only, and thus cannot be used for translation\nor generating text in other languages.", "#### Limitations and biases\n\n\nThe core functionality of GPT-NeoX-20B is to take a string of text and predict\nthe next token. Remember that the statistically most likely next token need\nnot result in the most “accurate” text. Never rely on GPT-NeoX-20B to produce\nfactually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nGPT-NeoX-20B may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nWe recommend curating the outputs of this model before presenting it to a human\nreader. Please inform your audience that you are using artificially generated\ntext.", "#### How to use\n\n\nIf you simply want to try out some prompts, check out this\nplayground.\n\n\nGPT-NeoX-20B can be loaded using the 'AutoModelForCausalLM' functionality:", "### Training", "#### Training dataset\n\n\nThe Pile is a 825GiB general-purpose dataset in English. It was created by\nEleutherAI specifically for training large language models. It contains texts\nfrom 22 diverse sources, roughly broken down into five categories: academic\nwriting (e.g. arXiv), internet (e.g. CommonCrawl), prose (e.g. Project\nGutenberg), dialogue (e.g. YouTube subtitles), and miscellaneous (e.g. GitHub,\nEnron Emails). See the Pile paper for\na breakdown of all data sources, methodology, and a discussion of ethical\nimplications. Consult the datasheet for\nmore detailed documentation about the Pile and its component datasets. The\nPile can be downloaded from the official website,\nor from a community mirror.\n\n\nThe Pile was not deduplicated before being used to train GPT-NeoX-20B.", "#### Training procedure\n\n\nGPT-NeoX-20B was trained with a batch size of approximately 3.15M tokens\n(1538 sequences of 2048 tokens each), for a total of 150,000 steps. Tensor\nparallelism and pipeline parallelism were used to distribute the model across\nGPUs. Additional details about the training procedure are in Section 3 of\nthe accompanying paper.", "### Evaluations\n\n\n\n\nZero-shot performance on selected natural language tasks.\n\nThis is a heavily abridged version of the evaluation results. Appendix D of the\nGPT-NeoX-20B paper compares more model\nsizes, and contains additional evaluations, including on: zero and five-shot\nnatural language tasks, zero and five-shot Basic Arithmetic and MATH,\nand zero-shot Hendrycks tasks.", "### BibTeX\n\n\nTo cite the GPT-NeoX-20B paper:\n\n\nOpen LLM Leaderboard Evaluation Results\n=======================================\n\n\nDetailed results can be found here" ]

[ "stable-baselines3", "LunarLander-v2", "deep-reinforcement-learning", "reinforcement-learning", "model-index", "region:us" ]

[]

[]

[ "# PPO Agent playing LunarLander-v2\nThis is a trained model of a PPO agent playing LunarLander-v2\nusing the stable-baselines3 library.", "## Usage (with Stable-baselines3)\nTODO: Add your code" ]

[ "TAGS\n#stable-baselines3 #LunarLander-v2 #deep-reinforcement-learning #reinforcement-learning #model-index #region-us \n", "# PPO Agent playing LunarLander-v2\nThis is a trained model of a PPO agent playing LunarLander-v2\nusing the stable-baselines3 library.", "## Usage (with Stable-baselines3)\nTODO: Add your code" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2104.09864", "arxiv:2204.04541", "arxiv:2306.02254", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[ "2104.09864", "2204.04541", "2306.02254" ]

[]

[ "### COPA (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### HellaSwag (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### BoolQ (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### SentiNeg (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### WiC (F1)\n\n\n\n<img src=\"URL width=\"800px\">\n\n\nLimitations and Biases\n----------------------\n\n\nPolyglot-Ko has been trained to optimize next token prediction. Language models such as this are often used for a wide variety of tasks and it is important to be aware of possible unexpected outcomes. For instance, Polyglot-Ko will not always return the most factual or accurate response but the most statistically likely one. In addition, Polyglot may produce socially unacceptable or offensive content. We recommend having a human curator or other filtering mechanism to censor sensitive content.\n\n\nand Related Information", "### BibTeX entry\n\n\nIf you find our work useful, please consider citing:", "### Licensing\n\n\nAll our models are licensed under the terms of the Apache License 2.0.", "### Acknowledgement\n\n\nThis project was made possible thanks to the computing resources from URL, and thanks to TUNiB for providing a large-scale Korean dataset for this work." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2104.09864 #arxiv-2204.04541 #arxiv-2306.02254 #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n", "### COPA (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### HellaSwag (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### BoolQ (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### SentiNeg (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### WiC (F1)\n\n\n\n<img src=\"URL width=\"800px\">\n\n\nLimitations and Biases\n----------------------\n\n\nPolyglot-Ko has been trained to optimize next token prediction. Language models such as this are often used for a wide variety of tasks and it is important to be aware of possible unexpected outcomes. For instance, Polyglot-Ko will not always return the most factual or accurate response but the most statistically likely one. In addition, Polyglot may produce socially unacceptable or offensive content. We recommend having a human curator or other filtering mechanism to censor sensitive content.\n\n\nand Related Information", "### BibTeX entry\n\n\nIf you find our work useful, please consider citing:", "### Licensing\n\n\nAll our models are licensed under the terms of the Apache License 2.0.", "### Acknowledgement\n\n\nThis project was made possible thanks to the computing resources from URL, and thanks to TUNiB for providing a large-scale Korean dataset for this work." ]

[ "transformers", "tensorboard", "safetensors", "xlm-roberta", "multiple-choice", "generated_from_trainer", "base_model:FacebookAI/xlm-roberta-large", "license:mit", "endpoints_compatible", "region:us" ]

[]

[]

[ "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 5e-05\n* train\\_batch\\_size: 16\n* eval\\_batch\\_size: 16\n* seed: 42\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* num\\_epochs: 10", "### Training results", "### Framework versions\n\n\n* Transformers 4.40.0\n* Pytorch 2.2.1+cu121\n* Datasets 2.19.0\n* Tokenizers 0.19.1" ]

[ "TAGS\n#transformers #tensorboard #safetensors #xlm-roberta #multiple-choice #generated_from_trainer #base_model-FacebookAI/xlm-roberta-large #license-mit #endpoints_compatible #region-us \n", "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 5e-05\n* train\\_batch\\_size: 16\n* eval\\_batch\\_size: 16\n* seed: 42\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* num\\_epochs: 10", "### Training results", "### Framework versions\n\n\n* Transformers 4.40.0\n* Pytorch 2.2.1+cu121\n* Datasets 2.19.0\n* Tokenizers 0.19.1" ]

[ "peft", "arxiv:1910.09700", "base_model:TinyLlama/TinyLlama-1.1B-Chat-v1.0", "region:us" ]

[ "1910.09700" ]

[]

[ "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\n\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact", "### Framework versions\n\n- PEFT 0.10.1.dev0" ]

[ "TAGS\n#peft #arxiv-1910.09700 #base_model-TinyLlama/TinyLlama-1.1B-Chat-v1.0 #region-us \n", "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\n\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact", "### Framework versions\n\n- PEFT 0.10.1.dev0" ]

[ "peft", "safetensors", "llama-factory", "lora", "generated_from_trainer", "base_model:happyrobot-ai/Llama3_function_calling_v1", "license:other", "region:us" ]

[]

[]

[ "# v3\n\nThis model is a fine-tuned version of happyrobot-ai/Llama3_function_calling_v1 on the jobandtalent_sft_v3 dataset.", "## Model description\n\nMore information needed", "## Intended uses & limitations\n\nMore information needed", "## Training and evaluation data\n\nMore information needed", "## Training procedure", "### Training hyperparameters\n\nThe following hyperparameters were used during training:\n- learning_rate: 0.0001\n- train_batch_size: 2\n- eval_batch_size: 8\n- seed: 42\n- distributed_type: multi-GPU\n- num_devices: 4\n- gradient_accumulation_steps: 2\n- total_train_batch_size: 16\n- total_eval_batch_size: 32\n- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n- lr_scheduler_type: cosine\n- num_epochs: 6.0", "### Training results", "### Framework versions\n\n- PEFT 0.10.0\n- Transformers 4.38.2\n- Pytorch 2.2.1+cu121\n- Datasets 2.18.0\n- Tokenizers 0.15.2" ]

[ "TAGS\n#peft #safetensors #llama-factory #lora #generated_from_trainer #base_model-happyrobot-ai/Llama3_function_calling_v1 #license-other #region-us \n", "# v3\n\nThis model is a fine-tuned version of happyrobot-ai/Llama3_function_calling_v1 on the jobandtalent_sft_v3 dataset.", "## Model description\n\nMore information needed", "## Intended uses & limitations\n\nMore information needed", "## Training and evaluation data\n\nMore information needed", "## Training procedure", "### Training hyperparameters\n\nThe following hyperparameters were used during training:\n- learning_rate: 0.0001\n- train_batch_size: 2\n- eval_batch_size: 8\n- seed: 42\n- distributed_type: multi-GPU\n- num_devices: 4\n- gradient_accumulation_steps: 2\n- total_train_batch_size: 16\n- total_eval_batch_size: 32\n- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n- lr_scheduler_type: cosine\n- num_epochs: 6.0", "### Training results", "### Framework versions\n\n- PEFT 0.10.0\n- Transformers 4.38.2\n- Pytorch 2.2.1+cu121\n- Datasets 2.18.0\n- Tokenizers 0.15.2" ]

[ "transformers", "safetensors", "mistral", "text-generation", "merge", "mergekit", "lazymergekit", "Weyaxi/Einstein-v6-7B", "argilla/CapybaraHermes-2.5-Mistral-7B", "conversational", "base_model:Weyaxi/Einstein-v6-7B", "base_model:argilla/CapybaraHermes-2.5-Mistral-7B", "license:apache-2.0", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "region:us" ]

[]

[]

[ "# J.O.S.I.E.3-Beta12-7B-slerp\n\nJ.O.S.I.E.3-Beta12-7B-slerp is a merge of the following models using LazyMergekit:\n* Weyaxi/Einstein-v6-7B\n* argilla/CapybaraHermes-2.5-Mistral-7B\n\nThis model has been further Finetuned on my custom J.O.S.I.E.v3.11 Dataset, in the ChatML prompt Format.", "## Configuration", "## Usage", "# Evaluation results:" ]

[ "TAGS\n#transformers #safetensors #mistral #text-generation #merge #mergekit #lazymergekit #Weyaxi/Einstein-v6-7B #argilla/CapybaraHermes-2.5-Mistral-7B #conversational #base_model-Weyaxi/Einstein-v6-7B #base_model-argilla/CapybaraHermes-2.5-Mistral-7B #license-apache-2.0 #autotrain_compatible #endpoints_compatible #text-generation-inference #region-us \n", "# J.O.S.I.E.3-Beta12-7B-slerp\n\nJ.O.S.I.E.3-Beta12-7B-slerp is a merge of the following models using LazyMergekit:\n* Weyaxi/Einstein-v6-7B\n* argilla/CapybaraHermes-2.5-Mistral-7B\n\nThis model has been further Finetuned on my custom J.O.S.I.E.v3.11 Dataset, in the ChatML prompt Format.", "## Configuration", "## Usage", "# Evaluation results:" ]

[ "diffusers", "safetensors", "arxiv:1910.09700", "endpoints_compatible", "diffusers:StableDiffusionXLPipeline", "region:us" ]

[ "1910.09700" ]

[]

[ "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a diffusers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "TAGS\n#diffusers #safetensors #arxiv-1910.09700 #endpoints_compatible #diffusers-StableDiffusionXLPipeline #region-us \n", "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a diffusers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "transformers", "tensorboard", "safetensors", "wav2vec2", "automatic-speech-recognition", "generated_from_trainer", "dataset:common_voice_11_0", "base_model:facebook/wav2vec2-xls-r-300m", "license:apache-2.0", "endpoints_compatible", "region:us" ]

[]

[]

[ "# wav2vec2-large-xls-r-300m-assamese\n\nThis model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice_11_0 dataset.", "## Model description\n\nMore information needed", "## Intended uses & limitations\n\nMore information needed", "## Training and evaluation data\n\nMore information needed", "## Training procedure", "### Training hyperparameters\n\nThe following hyperparameters were used during training:\n- learning_rate: 0.0003\n- train_batch_size: 16\n- eval_batch_size: 8\n- seed: 42\n- gradient_accumulation_steps: 2\n- total_train_batch_size: 32\n- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n- lr_scheduler_type: linear\n- lr_scheduler_warmup_steps: 500\n- num_epochs: 3\n- mixed_precision_training: Native AMP", "### Training results", "### Framework versions\n\n- Transformers 4.40.0\n- Pytorch 2.2.1+cu121\n- Datasets 2.19.0\n- Tokenizers 0.19.1" ]

[ "TAGS\n#transformers #tensorboard #safetensors #wav2vec2 #automatic-speech-recognition #generated_from_trainer #dataset-common_voice_11_0 #base_model-facebook/wav2vec2-xls-r-300m #license-apache-2.0 #endpoints_compatible #region-us \n", "# wav2vec2-large-xls-r-300m-assamese\n\nThis model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice_11_0 dataset.", "## Model description\n\nMore information needed", "## Intended uses & limitations\n\nMore information needed", "## Training and evaluation data\n\nMore information needed", "## Training procedure", "### Training hyperparameters\n\nThe following hyperparameters were used during training:\n- learning_rate: 0.0003\n- train_batch_size: 16\n- eval_batch_size: 8\n- seed: 42\n- gradient_accumulation_steps: 2\n- total_train_batch_size: 32\n- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n- lr_scheduler_type: linear\n- lr_scheduler_warmup_steps: 500\n- num_epochs: 3\n- mixed_precision_training: Native AMP", "### Training results", "### Framework versions\n\n- Transformers 4.40.0\n- Pytorch 2.2.1+cu121\n- Datasets 2.19.0\n- Tokenizers 0.19.1" ]

[ "transformers", "tensorboard", "safetensors", "vision-encoder-decoder", "generated_from_trainer", "dataset:imagefolder", "base_model:naver-clova-ix/donut-base", "license:mit", "endpoints_compatible", "region:us" ]

[]

[]

[ "# PolizzeDonut-UltimaProvaCluster-Cluster2di4-5epochs-Resol964x1350\n\nThis model is a fine-tuned version of naver-clova-ix/donut-base on the imagefolder dataset.", "## Model description\n\nMore information needed", "## Intended uses & limitations\n\nMore information needed", "## Training and evaluation data\n\nMore information needed", "## Training procedure", "### Training hyperparameters\n\nThe following hyperparameters were used during training:\n- learning_rate: 2e-05\n- train_batch_size: 2\n- eval_batch_size: 8\n- seed: 42\n- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n- lr_scheduler_type: linear\n- num_epochs: 5\n- mixed_precision_training: Native AMP", "### Training results", "### Framework versions\n\n- Transformers 4.38.2\n- Pytorch 2.2.2+cu121\n- Datasets 2.18.0\n- Tokenizers 0.15.2" ]

[ "TAGS\n#transformers #tensorboard #safetensors #vision-encoder-decoder #generated_from_trainer #dataset-imagefolder #base_model-naver-clova-ix/donut-base #license-mit #endpoints_compatible #region-us \n", "# PolizzeDonut-UltimaProvaCluster-Cluster2di4-5epochs-Resol964x1350\n\nThis model is a fine-tuned version of naver-clova-ix/donut-base on the imagefolder dataset.", "## Model description\n\nMore information needed", "## Intended uses & limitations\n\nMore information needed", "## Training and evaluation data\n\nMore information needed", "## Training procedure", "### Training hyperparameters\n\nThe following hyperparameters were used during training:\n- learning_rate: 2e-05\n- train_batch_size: 2\n- eval_batch_size: 8\n- seed: 42\n- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n- lr_scheduler_type: linear\n- num_epochs: 5\n- mixed_precision_training: Native AMP", "### Training results", "### Framework versions\n\n- Transformers 4.38.2\n- Pytorch 2.2.2+cu121\n- Datasets 2.18.0\n- Tokenizers 0.15.2" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2104.09864", "arxiv:2204.04541", "arxiv:2306.02254", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "4-bit", "region:us" ]

[ "2104.09864", "2204.04541", "2306.02254" ]

[]

[ "### COPA (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### HellaSwag (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### BoolQ (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### SentiNeg (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### WiC (F1)\n\n\n\n<img src=\"URL width=\"800px\">\n\n\nLimitations and Biases\n----------------------\n\n\nPolyglot-Ko has been trained to optimize next token prediction. Language models such as this are often used for a wide variety of tasks and it is important to be aware of possible unexpected outcomes. For instance, Polyglot-Ko will not always return the most factual or accurate response but the most statistically likely one. In addition, Polyglot may produce socially unacceptable or offensive content. We recommend having a human curator or other filtering mechanism to censor sensitive content.\n\n\nand Related Information", "### BibTeX entry\n\n\nIf you find our work useful, please consider citing:", "### Licensing\n\n\nAll our models are licensed under the terms of the Apache License 2.0.", "### Acknowledgement\n\n\nThis project was made possible thanks to the computing resources from URL, and thanks to TUNiB for providing a large-scale Korean dataset for this work." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2104.09864 #arxiv-2204.04541 #arxiv-2306.02254 #autotrain_compatible #endpoints_compatible #text-generation-inference #4-bit #region-us \n", "### COPA (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### HellaSwag (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### BoolQ (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### SentiNeg (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### WiC (F1)\n\n\n\n<img src=\"URL width=\"800px\">\n\n\nLimitations and Biases\n----------------------\n\n\nPolyglot-Ko has been trained to optimize next token prediction. Language models such as this are often used for a wide variety of tasks and it is important to be aware of possible unexpected outcomes. For instance, Polyglot-Ko will not always return the most factual or accurate response but the most statistically likely one. In addition, Polyglot may produce socially unacceptable or offensive content. We recommend having a human curator or other filtering mechanism to censor sensitive content.\n\n\nand Related Information", "### BibTeX entry\n\n\nIf you find our work useful, please consider citing:", "### Licensing\n\n\nAll our models are licensed under the terms of the Apache License 2.0.", "### Acknowledgement\n\n\nThis project was made possible thanks to the computing resources from URL, and thanks to TUNiB for providing a large-scale Korean dataset for this work." ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "4-bit", "region:us" ]

[ "2101.00027", "2201.07311" ]

[]

[ "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change over the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-160M for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-160M as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-160M has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-160M will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-160M to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-160M may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-160M.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-160M.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #4-bit #region-us \n", "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change over the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-160M for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-160M as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-160M has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-160M will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-160M to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-160M may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-160M.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-160M.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[ "2101.00027", "2201.07311" ]

[]

[ "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change over the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-160M for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-160M as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-160M has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-160M will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-160M to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-160M may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-160M.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-160M.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n", "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change over the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-160M for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-160M as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-160M has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-160M will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-160M to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-160M may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-160M.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-160M.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "vit", "image-classification", "generated_from_trainer", "base_model:google/vit-base-patch16-224-in21k", "license:apache-2.0", "autotrain_compatible", "endpoints_compatible", "region:us" ]

[]

[]

[ "# graph-classification-1\n\nThis model is a fine-tuned version of google/vit-base-patch16-224-in21k on an unknown dataset.\nIt achieves the following results on the evaluation set:\n- eval_loss: 0.2365\n- eval_accuracy: 0.9285\n- eval_runtime: 32.2134\n- eval_samples_per_second: 142.332\n- eval_steps_per_second: 4.47\n- epoch: 8.0\n- step: 1148", "## Model description\n\nMore information needed", "## Intended uses & limitations\n\nMore information needed", "## Training and evaluation data\n\nMore information needed", "## Training procedure", "### Training hyperparameters\n\nThe following hyperparameters were used during training:\n- learning_rate: 5e-05\n- train_batch_size: 32\n- eval_batch_size: 32\n- seed: 42\n- gradient_accumulation_steps: 4\n- total_train_batch_size: 128\n- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n- lr_scheduler_type: linear\n- lr_scheduler_warmup_ratio: 0.1\n- num_epochs: 10\n- mixed_precision_training: Native AMP", "### Framework versions\n\n- Transformers 4.40.0\n- Pytorch 2.2.1+cu121\n- Datasets 2.19.0\n- Tokenizers 0.19.1" ]

[ "TAGS\n#transformers #safetensors #vit #image-classification #generated_from_trainer #base_model-google/vit-base-patch16-224-in21k #license-apache-2.0 #autotrain_compatible #endpoints_compatible #region-us \n", "# graph-classification-1\n\nThis model is a fine-tuned version of google/vit-base-patch16-224-in21k on an unknown dataset.\nIt achieves the following results on the evaluation set:\n- eval_loss: 0.2365\n- eval_accuracy: 0.9285\n- eval_runtime: 32.2134\n- eval_samples_per_second: 142.332\n- eval_steps_per_second: 4.47\n- epoch: 8.0\n- step: 1148", "## Model description\n\nMore information needed", "## Intended uses & limitations\n\nMore information needed", "## Training and evaluation data\n\nMore information needed", "## Training procedure", "### Training hyperparameters\n\nThe following hyperparameters were used during training:\n- learning_rate: 5e-05\n- train_batch_size: 32\n- eval_batch_size: 32\n- seed: 42\n- gradient_accumulation_steps: 4\n- total_train_batch_size: 128\n- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n- lr_scheduler_type: linear\n- lr_scheduler_warmup_ratio: 0.1\n- num_epochs: 10\n- mixed_precision_training: Native AMP", "### Framework versions\n\n- Transformers 4.40.0\n- Pytorch 2.2.1+cu121\n- Datasets 2.19.0\n- Tokenizers 0.19.1" ]

[ "transformers", "tensorboard", "safetensors", "roberta", "token-classification", "generated_from_trainer", "base_model:vinai/phobert-base-v2", "autotrain_compatible", "endpoints_compatible", "region:us" ]

[]

[]

[ "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 1e-05\n* train\\_batch\\_size: 1\n* eval\\_batch\\_size: 1\n* seed: 42\n* gradient\\_accumulation\\_steps: 16\n* total\\_train\\_batch\\_size: 16\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* num\\_epochs: 5", "### Training results", "### Framework versions\n\n\n* Transformers 4.39.0.dev0\n* Pytorch 2.2.1+cu118\n* Datasets 2.18.0\n* Tokenizers 0.15.2" ]

[ "TAGS\n#transformers #tensorboard #safetensors #roberta #token-classification #generated_from_trainer #base_model-vinai/phobert-base-v2 #autotrain_compatible #endpoints_compatible #region-us \n", "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 1e-05\n* train\\_batch\\_size: 1\n* eval\\_batch\\_size: 1\n* seed: 42\n* gradient\\_accumulation\\_steps: 16\n* total\\_train\\_batch\\_size: 16\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* num\\_epochs: 5", "### Training results", "### Framework versions\n\n\n* Transformers 4.39.0.dev0\n* Pytorch 2.2.1+cu118\n* Datasets 2.18.0\n* Tokenizers 0.15.2" ]

[ "peft", "tensorboard", "safetensors", "generated_from_trainer", "base_model:cardiffnlp/twitter-roberta-base-sentiment-latest", "has_space", "region:us" ]

[]

[]

[ "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 0.001\n* train\\_batch\\_size: 4\n* eval\\_batch\\_size: 4\n* seed: 42\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* num\\_epochs: 50", "### Training results", "### Framework versions\n\n\n* PEFT 0.10.0\n* Transformers 4.38.2\n* Pytorch 2.2.1\n* Datasets 2.18.0\n* Tokenizers 0.15.2" ]

[ "TAGS\n#peft #tensorboard #safetensors #generated_from_trainer #base_model-cardiffnlp/twitter-roberta-base-sentiment-latest #has_space #region-us \n", "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 0.001\n* train\\_batch\\_size: 4\n* eval\\_batch\\_size: 4\n* seed: 42\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* num\\_epochs: 50", "### Training results", "### Framework versions\n\n\n* PEFT 0.10.0\n* Transformers 4.38.2\n* Pytorch 2.2.1\n* Datasets 2.18.0\n* Tokenizers 0.15.2" ]

[ "peft", "safetensors", "generated_from_trainer", "base_model:microsoft/phi-2", "license:mit", "region:us" ]

[]

[]

[ "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 2e-05\n* train\\_batch\\_size: 2\n* eval\\_batch\\_size: 8\n* seed: 42\n* gradient\\_accumulation\\_steps: 6\n* total\\_train\\_batch\\_size: 12\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* num\\_epochs: 4", "### Training results", "### Framework versions\n\n\n* PEFT 0.10.0\n* Transformers 4.40.0\n* Pytorch 2.2.2+cu121\n* Datasets 2.19.0\n* Tokenizers 0.19.1" ]

[ "TAGS\n#peft #safetensors #generated_from_trainer #base_model-microsoft/phi-2 #license-mit #region-us \n", "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 2e-05\n* train\\_batch\\_size: 2\n* eval\\_batch\\_size: 8\n* seed: 42\n* gradient\\_accumulation\\_steps: 6\n* total\\_train\\_batch\\_size: 12\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* num\\_epochs: 4", "### Training results", "### Framework versions\n\n\n* PEFT 0.10.0\n* Transformers 4.40.0\n* Pytorch 2.2.2+cu121\n* Datasets 2.19.0\n* Tokenizers 0.19.1" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "4-bit", "region:us" ]

[ "2101.00027", "2201.07311" ]

[]

[ "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change over the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-1B for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-1B as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-1B has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-1B will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-1B to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-1B may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-1B.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-1B.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #4-bit #region-us \n", "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change over the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-1B for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-1B as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-1B has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-1B will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-1B to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-1B may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-1B.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-1B.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "arxiv:1910.09700", "endpoints_compatible", "region:us" ]

[ "1910.09700" ]

[]

[ "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a transformers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "TAGS\n#transformers #safetensors #arxiv-1910.09700 #endpoints_compatible #region-us \n", "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a transformers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[ "2101.00027", "2201.07311" ]

[]

[ "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change over the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-1B for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-1B as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-1B has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-1B will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-1B to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-1B may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-1B.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-1B.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n", "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change over the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-1B for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-1B as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-1B has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-1B will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-1B to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-1B may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-1B.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-1B.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "4-bit", "region:us" ]

[ "2101.00027", "2201.07311" ]

[]

[ "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change over the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-70M for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-70M as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-70M has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-70M will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-70M to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-70M may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-70M.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-70M.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #4-bit #region-us \n", "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change over the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-70M for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-70M as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-70M has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-70M will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-70M to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-70M may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-70M.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-70M.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "tensorboard", "safetensors", "albert", "text-classification", "generated_from_trainer", "base_model:albert/albert-base-v2", "license:apache-2.0", "autotrain_compatible", "endpoints_compatible", "region:us" ]

[]

[]

[ "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 2e-05\n* train\\_batch\\_size: 16\n* eval\\_batch\\_size: 16\n* seed: 42\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* num\\_epochs: 2", "### Training results", "### Framework versions\n\n\n* Transformers 4.40.0\n* Pytorch 2.2.1+cu121\n* Datasets 2.19.0\n* Tokenizers 0.19.1" ]

[ "TAGS\n#transformers #tensorboard #safetensors #albert #text-classification #generated_from_trainer #base_model-albert/albert-base-v2 #license-apache-2.0 #autotrain_compatible #endpoints_compatible #region-us \n", "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 2e-05\n* train\\_batch\\_size: 16\n* eval\\_batch\\_size: 16\n* seed: 42\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* num\\_epochs: 2", "### Training results", "### Framework versions\n\n\n* Transformers 4.40.0\n* Pytorch 2.2.1+cu121\n* Datasets 2.19.0\n* Tokenizers 0.19.1" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2104.09864", "arxiv:2204.04541", "arxiv:2306.02254", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[ "2104.09864", "2204.04541", "2306.02254" ]

[]

[ "### COPA (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### HellaSwag (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### BoolQ (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### SentiNeg (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### WiC (F1)\n\n\n\n<img src=\"URL width=\"800px\">\n\n\nLimitations and Biases\n----------------------\n\n\nPolyglot-Ko has been trained to optimize next token prediction. Language models such as this are often used for a wide variety of tasks and it is important to be aware of possible unexpected outcomes. For instance, Polyglot-Ko will not always return the most factual or accurate response but the most statistically likely one. In addition, Polyglot may produce socially unacceptable or offensive content. We recommend having a human curator or other filtering mechanism to censor sensitive content.\n\n\nand Related Information", "### BibTeX entry\n\n\nIf you find our work useful, please consider citing:", "### Licensing\n\n\nAll our models are licensed under the terms of the Apache License 2.0.", "### Acknowledgement\n\n\nThis project was made possible thanks to the computing resources from URL, and thanks to TUNiB for providing a large-scale Korean dataset for this work." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2104.09864 #arxiv-2204.04541 #arxiv-2306.02254 #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n", "### COPA (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### HellaSwag (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### BoolQ (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### SentiNeg (F1)\n\n\n\n<img src=\"URL width=\"800px\">", "### WiC (F1)\n\n\n\n<img src=\"URL width=\"800px\">\n\n\nLimitations and Biases\n----------------------\n\n\nPolyglot-Ko has been trained to optimize next token prediction. Language models such as this are often used for a wide variety of tasks and it is important to be aware of possible unexpected outcomes. For instance, Polyglot-Ko will not always return the most factual or accurate response but the most statistically likely one. In addition, Polyglot may produce socially unacceptable or offensive content. We recommend having a human curator or other filtering mechanism to censor sensitive content.\n\n\nand Related Information", "### BibTeX entry\n\n\nIf you find our work useful, please consider citing:", "### Licensing\n\n\nAll our models are licensed under the terms of the Apache License 2.0.", "### Acknowledgement\n\n\nThis project was made possible thanks to the computing resources from URL, and thanks to TUNiB for providing a large-scale Korean dataset for this work." ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[ "2101.00027", "2201.07311" ]

[]

[ "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change over the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-70M for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-70M as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-70M has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-70M will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-70M to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-70M may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-70M.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-70M.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n", "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change over the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-70M for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-70M as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-70M has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-70M will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-70M to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-70M may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-70M.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-70M.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "4-bit", "region:us" ]

[ "2101.00027", "2201.07311" ]

[]

[ "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change over the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-410M for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-410M as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-410M has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-410M will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-410M to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-410M may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-410M.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-410M.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #4-bit #region-us \n", "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change over the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-410M for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-410M as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-410M has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-410M will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-410M to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-410M may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-410M.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-410M.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "mistral", "text-generation", "arxiv:1910.09700", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "region:us" ]

[ "1910.09700" ]

[]

[ "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a transformers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "TAGS\n#transformers #safetensors #mistral #text-generation #arxiv-1910.09700 #autotrain_compatible #endpoints_compatible #text-generation-inference #region-us \n", "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a transformers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[ "2101.00027", "2201.07311" ]

[]

[ "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change over the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-410M for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-410M as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-410M has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-410M will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-410M to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-410M may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-410M.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-410M.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n", "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change over the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-410M for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-410M as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-410M has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-410M will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-410M to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-410M may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-410M.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-410M.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "4-bit", "region:us" ]

[ "2101.00027", "2201.07311" ]

[]

[ "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-160M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-160M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-160M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-160M-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-160M-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-160M-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-160M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-160M-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #4-bit #region-us \n", "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-160M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-160M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-160M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-160M-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-160M-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-160M-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-160M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-160M-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[ "2101.00027", "2201.07311" ]

[]

[ "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-160M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-160M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-160M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-160M-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-160M-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-160M-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-160M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-160M-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n", "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-160M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-160M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-160M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-160M-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-160M-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-160M-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-160M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-160M-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "gguf", "llama", "text-generation", "not-for-all-audiences", "art", "en", "dataset:KBlueLeaf/danbooru2023-sqlite", "license:cc-by-nc-sa-4.0", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "region:us" ]

[]

[ "en" ]

[ "# DanTagGen - delta\nDanTagGen(Danbooru Tag Generator) is inspired from p1atdev's dart project.\nBut with different arch, dataset, format and different training strategy.", "## Difference between versions\nalpha: pretrain on 2M dataset, smaller batch size. Limited ability\nbeta: pretrain on 5.3M dataset, larger batch size. More stable, better ability with only a few information provided.\ndelta: pretrain on 7.2M dataset, larger batch size. Slightly underfit but better diversity. quality tag introduced.", "## Model arch\nThis version of DTG is trained from scratch with 400M param LLaMA arch.(In my personal preference I will call it NanoLLaMA)\nSince it is llama arch. Theoritically it should be able to be used in any LLaMA inference interface.\n\nThis repo also provided converted FP16 gguf model and quantized 8bit/6bit gguf models.\nBasically it is recommended to use URL or llama-cpp-python to run this model. Which will be very fast.", "## Format\n\n\nfor example:\n\n\nAnd you may get something like:", "## Dataset and Training\nI use the trainer I implemented in HakuPhi to run the training.\nwith 10epoch on 7.2M data. This model have roughly 10~15B token seen.\n\nThe dataset is exported by HakuBooru with my danbooru sqlite database. Use the percentile of fav_count on each rating to filter the data. (2M = top 25%, 5.3M = top 75%)", "## Utilities\nHF space: URL\nDemo for DTG + Kohaku XL Epsilon: URL\nSD-WebUI Extension: URL\nComfyUI Node: URL" ]

[ "TAGS\n#transformers #safetensors #gguf #llama #text-generation #not-for-all-audiences #art #en #dataset-KBlueLeaf/danbooru2023-sqlite #license-cc-by-nc-sa-4.0 #autotrain_compatible #endpoints_compatible #text-generation-inference #region-us \n", "# DanTagGen - delta\nDanTagGen(Danbooru Tag Generator) is inspired from p1atdev's dart project.\nBut with different arch, dataset, format and different training strategy.", "## Difference between versions\nalpha: pretrain on 2M dataset, smaller batch size. Limited ability\nbeta: pretrain on 5.3M dataset, larger batch size. More stable, better ability with only a few information provided.\ndelta: pretrain on 7.2M dataset, larger batch size. Slightly underfit but better diversity. quality tag introduced.", "## Model arch\nThis version of DTG is trained from scratch with 400M param LLaMA arch.(In my personal preference I will call it NanoLLaMA)\nSince it is llama arch. Theoritically it should be able to be used in any LLaMA inference interface.\n\nThis repo also provided converted FP16 gguf model and quantized 8bit/6bit gguf models.\nBasically it is recommended to use URL or llama-cpp-python to run this model. Which will be very fast.", "## Format\n\n\nfor example:\n\n\nAnd you may get something like:", "## Dataset and Training\nI use the trainer I implemented in HakuPhi to run the training.\nwith 10epoch on 7.2M data. This model have roughly 10~15B token seen.\n\nThe dataset is exported by HakuBooru with my danbooru sqlite database. Use the percentile of fav_count on each rating to filter the data. (2M = top 25%, 5.3M = top 75%)", "## Utilities\nHF space: URL\nDemo for DTG + Kohaku XL Epsilon: URL\nSD-WebUI Extension: URL\nComfyUI Node: URL" ]

[ "transformers", "safetensors", "arxiv:1910.09700", "endpoints_compatible", "region:us" ]

[ "1910.09700" ]

[]

[ "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a transformers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "TAGS\n#transformers #safetensors #arxiv-1910.09700 #endpoints_compatible #region-us \n", "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a transformers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "transformers", "tensorboard", "safetensors", "longformer", "question-answering", "generated_from_trainer", "base_model:valhalla/longformer-base-4096-finetuned-squadv1", "license:mit", "endpoints_compatible", "region:us" ]

[]

[]

[ "# LongformerTest\n\nThis model is a fine-tuned version of valhalla/longformer-base-4096-finetuned-squadv1 on an unknown dataset.", "## Model description\n\nMore information needed", "## Intended uses & limitations\n\nMore information needed", "## Training and evaluation data\n\nMore information needed", "## Training procedure", "### Training hyperparameters\n\nThe following hyperparameters were used during training:\n- learning_rate: 2e-05\n- train_batch_size: 2\n- eval_batch_size: 0\n- seed: 42\n- gradient_accumulation_steps: 4\n- total_train_batch_size: 8\n- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n- lr_scheduler_type: linear\n- num_epochs: 3", "### Training results", "### Framework versions\n\n- Transformers 4.40.0\n- Pytorch 2.2.1+cu121\n- Datasets 2.19.0\n- Tokenizers 0.19.1" ]

[ "TAGS\n#transformers #tensorboard #safetensors #longformer #question-answering #generated_from_trainer #base_model-valhalla/longformer-base-4096-finetuned-squadv1 #license-mit #endpoints_compatible #region-us \n", "# LongformerTest\n\nThis model is a fine-tuned version of valhalla/longformer-base-4096-finetuned-squadv1 on an unknown dataset.", "## Model description\n\nMore information needed", "## Intended uses & limitations\n\nMore information needed", "## Training and evaluation data\n\nMore information needed", "## Training procedure", "### Training hyperparameters\n\nThe following hyperparameters were used during training:\n- learning_rate: 2e-05\n- train_batch_size: 2\n- eval_batch_size: 0\n- seed: 42\n- gradient_accumulation_steps: 4\n- total_train_batch_size: 8\n- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n- lr_scheduler_type: linear\n- num_epochs: 3", "### Training results", "### Framework versions\n\n- Transformers 4.40.0\n- Pytorch 2.2.1+cu121\n- Datasets 2.19.0\n- Tokenizers 0.19.1" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "4-bit", "region:us" ]

[ "2101.00027", "2201.07311" ]

[]

[ "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-1.4B-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-1.4B-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-1.4B-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-1.4B-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-1.4B-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-1.4B-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-1.4B-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-1.4B-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #4-bit #region-us \n", "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-1.4B-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-1.4B-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-1.4B-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-1.4B-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-1.4B-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-1.4B-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-1.4B-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-1.4B-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[ "2101.00027", "2201.07311" ]

[]

[ "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-1.4B-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-1.4B-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-1.4B-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-1.4B-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-1.4B-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-1.4B-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-1.4B-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-1.4B-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n", "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-1.4B-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-1.4B-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-1.4B-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-1.4B-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-1.4B-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-1.4B-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-1.4B-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-1.4B-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "4-bit", "region:us" ]

[ "2101.00027", "2201.07311" ]

[]

[ "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-1B-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-1B-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-1B-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-1B-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-1B-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-1B-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-1B-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-1B-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #4-bit #region-us \n", "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-1B-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-1B-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-1B-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-1B-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-1B-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-1B-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-1B-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-1B-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "license:apache-2.0", "region:us" ]

[]

[]

[]

[ "TAGS\n#license-apache-2.0 #region-us \n" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[ "2101.00027", "2201.07311" ]

[]

[ "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-1B-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-1B-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-1B-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-1B-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-1B-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-1B-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-1B-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-1B-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n", "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-1B-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-1B-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-1B-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-1B-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-1B-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-1B-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-1B-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-1B-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "diffusers", "safetensors", "text-to-image", "stable-diffusion", "license:creativeml-openrail-m", "endpoints_compatible", "diffusers:StableDiffusionPipeline", "region:us" ]

[]

[]

[ "### model Dreambooth model trained by amilyjenksy with TheLastBen's fast-DreamBooth notebook\n\n\nTest the concept via A1111 Colab fast-Colab-A1111\n\nSample pictures of this concept:" ]

[ "TAGS\n#diffusers #safetensors #text-to-image #stable-diffusion #license-creativeml-openrail-m #endpoints_compatible #diffusers-StableDiffusionPipeline #region-us \n", "### model Dreambooth model trained by amilyjenksy with TheLastBen's fast-DreamBooth notebook\n\n\nTest the concept via A1111 Colab fast-Colab-A1111\n\nSample pictures of this concept:" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "4-bit", "region:us" ]

[ "2101.00027", "2201.07311" ]

[]

[ "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-70M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-70M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-70M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-70M-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-70M-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-70M-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-70M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-70M-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #4-bit #region-us \n", "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-70M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-70M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-70M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-70M-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-70M-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-70M-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-70M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-70M-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "peft", "arxiv:1910.09700", "base_model:TinyLlama/TinyLlama-1.1B-Chat-v1.0", "region:us" ]

[ "1910.09700" ]

[]

[ "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\n\n\n- Developed by: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact", "## Training procedure\n\n\nThe following 'bitsandbytes' quantization config was used during training:\n- load_in_8bit: False\n- load_in_4bit: True\n- llm_int8_threshold: 6.0\n- llm_int8_skip_modules: None\n- llm_int8_enable_fp32_cpu_offload: False\n- llm_int8_has_fp16_weight: False\n- bnb_4bit_quant_type: nf4\n- bnb_4bit_use_double_quant: True\n- bnb_4bit_compute_dtype: bfloat16", "### Framework versions\n\n\n- PEFT 0.7.0.dev0", "## Training procedure\n\n\nThe following 'bitsandbytes' quantization config was used during training:\n- load_in_8bit: False\n- load_in_4bit: True\n- llm_int8_threshold: 6.0\n- llm_int8_skip_modules: None\n- llm_int8_enable_fp32_cpu_offload: False\n- llm_int8_has_fp16_weight: False\n- bnb_4bit_quant_type: nf4\n- bnb_4bit_use_double_quant: True\n- bnb_4bit_compute_dtype: bfloat16", "### Framework versions\n\n\n- PEFT 0.7.0.dev0" ]

[ "TAGS\n#peft #arxiv-1910.09700 #base_model-TinyLlama/TinyLlama-1.1B-Chat-v1.0 #region-us \n", "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\n\n\n- Developed by: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact", "## Training procedure\n\n\nThe following 'bitsandbytes' quantization config was used during training:\n- load_in_8bit: False\n- load_in_4bit: True\n- llm_int8_threshold: 6.0\n- llm_int8_skip_modules: None\n- llm_int8_enable_fp32_cpu_offload: False\n- llm_int8_has_fp16_weight: False\n- bnb_4bit_quant_type: nf4\n- bnb_4bit_use_double_quant: True\n- bnb_4bit_compute_dtype: bfloat16", "### Framework versions\n\n\n- PEFT 0.7.0.dev0", "## Training procedure\n\n\nThe following 'bitsandbytes' quantization config was used during training:\n- load_in_8bit: False\n- load_in_4bit: True\n- llm_int8_threshold: 6.0\n- llm_int8_skip_modules: None\n- llm_int8_enable_fp32_cpu_offload: False\n- llm_int8_has_fp16_weight: False\n- bnb_4bit_quant_type: nf4\n- bnb_4bit_use_double_quant: True\n- bnb_4bit_compute_dtype: bfloat16", "### Framework versions\n\n\n- PEFT 0.7.0.dev0" ]

[ "peft", "arxiv:1910.09700", "base_model:TinyLlama/TinyLlama-1.1B-Chat-v1.0", "region:us" ]

[ "1910.09700" ]

[]

[ "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\n\n\n- Developed by: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact", "## Training procedure\n\n\nThe following 'bitsandbytes' quantization config was used during training:\n- load_in_8bit: False\n- load_in_4bit: True\n- llm_int8_threshold: 6.0\n- llm_int8_skip_modules: None\n- llm_int8_enable_fp32_cpu_offload: False\n- llm_int8_has_fp16_weight: False\n- bnb_4bit_quant_type: nf4\n- bnb_4bit_use_double_quant: True\n- bnb_4bit_compute_dtype: bfloat16", "### Framework versions\n\n\n- PEFT 0.7.0.dev0" ]

[ "TAGS\n#peft #arxiv-1910.09700 #base_model-TinyLlama/TinyLlama-1.1B-Chat-v1.0 #region-us \n", "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\n\n\n- Developed by: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact", "## Training procedure\n\n\nThe following 'bitsandbytes' quantization config was used during training:\n- load_in_8bit: False\n- load_in_4bit: True\n- llm_int8_threshold: 6.0\n- llm_int8_skip_modules: None\n- llm_int8_enable_fp32_cpu_offload: False\n- llm_int8_has_fp16_weight: False\n- bnb_4bit_quant_type: nf4\n- bnb_4bit_use_double_quant: True\n- bnb_4bit_compute_dtype: bfloat16", "### Framework versions\n\n\n- PEFT 0.7.0.dev0" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[ "2101.00027", "2201.07311" ]

[]

[ "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-70M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-70M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-70M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-70M-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-70M-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-70M-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-70M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-70M-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n", "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-70M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-70M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-70M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-70M-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-70M-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-70M-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-70M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-70M-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "4-bit", "region:us" ]

[ "2101.00027", "2201.07311" ]

[]

[ "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-410M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-410M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-410M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-410M-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-410M-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-410M-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-410M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-410M-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #4-bit #region-us \n", "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-410M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-410M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-410M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-410M-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-410M-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-410M-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-410M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-410M-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[ "2101.00027", "2201.07311" ]

[]

[ "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-410M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-410M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-410M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-410M-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-410M-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-410M-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-410M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-410M-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n", "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-410M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-410M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-410M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-410M-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-410M-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-410M-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-410M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-410M-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "llama", "text-generation", "conversational", "arxiv:1910.09700", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "region:us" ]

[ "1910.09700" ]

[]

[ "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a transformers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "TAGS\n#transformers #safetensors #llama #text-generation #conversational #arxiv-1910.09700 #autotrain_compatible #endpoints_compatible #text-generation-inference #region-us \n", "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a transformers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "transformers", "tensorboard", "safetensors", "albert", "multiple-choice", "generated_from_trainer", "base_model:albert/albert-base-v2", "license:apache-2.0", "endpoints_compatible", "region:us" ]

[]

[]

[ "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 5e-05\n* train\\_batch\\_size: 16\n* eval\\_batch\\_size: 16\n* seed: 42\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* num\\_epochs: 10", "### Training results", "### Framework versions\n\n\n* Transformers 4.40.0\n* Pytorch 2.2.1+cu121\n* Datasets 2.19.0\n* Tokenizers 0.19.1" ]

[ "TAGS\n#transformers #tensorboard #safetensors #albert #multiple-choice #generated_from_trainer #base_model-albert/albert-base-v2 #license-apache-2.0 #endpoints_compatible #region-us \n", "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 5e-05\n* train\\_batch\\_size: 16\n* eval\\_batch\\_size: 16\n* seed: 42\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* num\\_epochs: 10", "### Training results", "### Framework versions\n\n\n* Transformers 4.40.0\n* Pytorch 2.2.1+cu121\n* Datasets 2.19.0\n* Tokenizers 0.19.1" ]

[ "diffusers", "safetensors", "arxiv:1910.09700", "endpoints_compatible", "diffusers:StableDiffusionXLPipeline", "region:us" ]

[ "1910.09700" ]

[]

[ "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a diffusers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "TAGS\n#diffusers #safetensors #arxiv-1910.09700 #endpoints_compatible #diffusers-StableDiffusionXLPipeline #region-us \n", "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a diffusers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "peft", "safetensors", "generated_from_trainer", "base_model:TheBloke/Llama-2-7B-fp16", "region:us" ]

[]

[]

[ "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 0.0003\n* train\\_batch\\_size: 2\n* eval\\_batch\\_size: 8\n* seed: 42\n* gradient\\_accumulation\\_steps: 10\n* total\\_train\\_batch\\_size: 20\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* training\\_steps: 50", "### Training results", "### Framework versions\n\n\n* PEFT 0.10.0\n* Transformers 4.39.3\n* Pytorch 2.2.2+cu121\n* Datasets 2.18.0\n* Tokenizers 0.15.2" ]

[ "TAGS\n#peft #safetensors #generated_from_trainer #base_model-TheBloke/Llama-2-7B-fp16 #region-us \n", "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 0.0003\n* train\\_batch\\_size: 2\n* eval\\_batch\\_size: 8\n* seed: 42\n* gradient\\_accumulation\\_steps: 10\n* total\\_train\\_batch\\_size: 20\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* training\\_steps: 50", "### Training results", "### Framework versions\n\n\n* PEFT 0.10.0\n* Transformers 4.39.3\n* Pytorch 2.2.2+cu121\n* Datasets 2.18.0\n* Tokenizers 0.15.2" ]

[ "transformers", "tensorboard", "safetensors", "t5", "text2text-generation", "generated_from_trainer", "base_model:google-t5/t5-small", "license:apache-2.0", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "region:us" ]

[]

[]

[ "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 2e-05\n* train\\_batch\\_size: 16\n* eval\\_batch\\_size: 16\n* seed: 42\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* num\\_epochs: 10\n* mixed\\_precision\\_training: Native AMP", "### Training results", "### Framework versions\n\n\n* Transformers 4.39.3\n* Pytorch 2.1.2\n* Datasets 2.18.0\n* Tokenizers 0.15.2" ]

[ "TAGS\n#transformers #tensorboard #safetensors #t5 #text2text-generation #generated_from_trainer #base_model-google-t5/t5-small #license-apache-2.0 #autotrain_compatible #endpoints_compatible #text-generation-inference #region-us \n", "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 2e-05\n* train\\_batch\\_size: 16\n* eval\\_batch\\_size: 16\n* seed: 42\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* num\\_epochs: 10\n* mixed\\_precision\\_training: Native AMP", "### Training results", "### Framework versions\n\n\n* Transformers 4.39.3\n* Pytorch 2.1.2\n* Datasets 2.18.0\n* Tokenizers 0.15.2" ]

[ "transformers", "pytorch", "mistral", "text-generation", "text-generation-inference", "unsloth", "trl", "sft", "conversational", "en", "base_model:unsloth/mistral-7b-instruct-v0.2-bnb-4bit", "license:apache-2.0", "autotrain_compatible", "endpoints_compatible", "region:us" ]

[]

[ "en" ]

[ "# Uploaded model\n\n- Developed by: shubham11\n- License: apache-2.0\n- Finetuned from model : unsloth/mistral-7b-instruct-v0.2-bnb-4bit\n\nThis mistral model was trained 2x faster with Unsloth and Huggingface's TRL library.\n\n<img src=\"URL width=\"200\"/>" ]

[ "TAGS\n#transformers #pytorch #mistral #text-generation #text-generation-inference #unsloth #trl #sft #conversational #en #base_model-unsloth/mistral-7b-instruct-v0.2-bnb-4bit #license-apache-2.0 #autotrain_compatible #endpoints_compatible #region-us \n", "# Uploaded model\n\n- Developed by: shubham11\n- License: apache-2.0\n- Finetuned from model : unsloth/mistral-7b-instruct-v0.2-bnb-4bit\n\nThis mistral model was trained 2x faster with Unsloth and Huggingface's TRL library.\n\n<img src=\"URL width=\"200\"/>" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "4-bit", "region:us" ]

[ "2101.00027", "2201.07311" ]

[]

[ "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change over the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-2.8B for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-2.8B as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-2.8B has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-2.8B will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-2.8B to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-2.8B may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-2.8B.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-2.8B.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #4-bit #region-us \n", "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change over the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-2.8B for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-2.8B as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-2.8B has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-2.8B will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-2.8B to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-2.8B may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-2.8B.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-2.8B.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2304.01373", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "4-bit", "region:us" ]

[ "2304.01373", "2101.00027", "2201.07311" ]

[]

[ "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-160M for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-160M as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-160M has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-160M will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-160M to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-160M may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-160M.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-160M.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2304.01373 #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #4-bit #region-us \n", "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-160M for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-160M as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-160M has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-160M will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-160M to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-160M may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-160M.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-160M.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "gguf", "en", "base_model:BAAI/JudgeLM-33B-v1.0", "endpoints_compatible", "region:us" ]

[]

[ "en" ]

[]

[ "TAGS\n#transformers #gguf #en #base_model-BAAI/JudgeLM-33B-v1.0 #endpoints_compatible #region-us \n" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2304.01373", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[ "2304.01373", "2101.00027", "2201.07311" ]

[]

[ "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-160M for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-160M as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-160M has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-160M will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-160M to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-160M may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-160M.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-160M.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2304.01373 #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n", "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-160M for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-160M as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-160M has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-160M will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-160M to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-160M may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-160M.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-160M.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "tf", "distilbert", "question-answering", "generated_from_keras_callback", "base_model:distilbert/distilbert-base-uncased", "license:apache-2.0", "endpoints_compatible", "region:us" ]

[]

[]

[ "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* optimizer: {'name': 'Adam', 'weight\\_decay': None, 'clipnorm': None, 'global\\_clipnorm': None, 'clipvalue': None, 'use\\_ema': False, 'ema\\_momentum': 0.99, 'ema\\_overwrite\\_frequency': None, 'jit\\_compile': True, 'is\\_legacy\\_optimizer': False, 'learning\\_rate': {'module': 'keras.optimizers.schedules', 'class\\_name': 'PolynomialDecay', 'config': {'initial\\_learning\\_rate': 2e-05, 'decay\\_steps': 60, 'end\\_learning\\_rate': 0.0, 'power': 1.0, 'cycle': False, 'name': None}, 'registered\\_name': None}, 'beta\\_1': 0.9, 'beta\\_2': 0.999, 'epsilon': 1e-08, 'amsgrad': False}\n* training\\_precision: float32", "### Training results", "### Framework versions\n\n\n* Transformers 4.40.0\n* TensorFlow 2.15.0\n* Datasets 2.19.0\n* Tokenizers 0.19.1" ]

[ "TAGS\n#transformers #tf #distilbert #question-answering #generated_from_keras_callback #base_model-distilbert/distilbert-base-uncased #license-apache-2.0 #endpoints_compatible #region-us \n", "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* optimizer: {'name': 'Adam', 'weight\\_decay': None, 'clipnorm': None, 'global\\_clipnorm': None, 'clipvalue': None, 'use\\_ema': False, 'ema\\_momentum': 0.99, 'ema\\_overwrite\\_frequency': None, 'jit\\_compile': True, 'is\\_legacy\\_optimizer': False, 'learning\\_rate': {'module': 'keras.optimizers.schedules', 'class\\_name': 'PolynomialDecay', 'config': {'initial\\_learning\\_rate': 2e-05, 'decay\\_steps': 60, 'end\\_learning\\_rate': 0.0, 'power': 1.0, 'cycle': False, 'name': None}, 'registered\\_name': None}, 'beta\\_1': 0.9, 'beta\\_2': 0.999, 'epsilon': 1e-08, 'amsgrad': False}\n* training\\_precision: float32", "### Training results", "### Framework versions\n\n\n* Transformers 4.40.0\n* TensorFlow 2.15.0\n* Datasets 2.19.0\n* Tokenizers 0.19.1" ]

[ "transformers", "safetensors", "xlm-roberta", "token-classification", "arxiv:1910.09700", "autotrain_compatible", "endpoints_compatible", "region:us" ]

[ "1910.09700" ]

[]

[ "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a transformers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "TAGS\n#transformers #safetensors #xlm-roberta #token-classification #arxiv-1910.09700 #autotrain_compatible #endpoints_compatible #region-us \n", "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a transformers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[ "2101.00027", "2201.07311" ]

[]

[ "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change over the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-2.8B for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-2.8B as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-2.8B has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-2.8B will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-2.8B to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-2.8B may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-2.8B.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-2.8B.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n", "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change over the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-2.8B for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-2.8B as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-2.8B has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-2.8B will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-2.8B to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-2.8B may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-2.8B.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-2.8B.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "4-bit", "region:us" ]

[ "2101.00027", "2201.07311" ]

[]

[ "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-2.8B-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-2.8B-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-2.8B-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-2.8B-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-2.8B-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-2.8B-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-2.8B-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-2.8B-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #4-bit #region-us \n", "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-2.8B-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-2.8B-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-2.8B-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-2.8B-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-2.8B-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-2.8B-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-2.8B-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-2.8B-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2304.01373", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "4-bit", "region:us" ]

[ "2304.01373", "2101.00027", "2201.07311" ]

[]

[ "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-160M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-160M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-160M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-160M-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-160M-deduped to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-160M-deduped may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-160M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nPythia-160M-deduped was trained on the Pile after the dataset has been globally\ndeduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2304.01373 #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #4-bit #region-us \n", "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-160M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-160M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-160M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-160M-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-160M-deduped to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-160M-deduped may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-160M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nPythia-160M-deduped was trained on the Pile after the dataset has been globally\ndeduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "text-generation-inference", "unsloth", "llama", "trl", "en", "base_model:unsloth/llama-3-8b-bnb-4bit", "license:apache-2.0", "endpoints_compatible", "region:us" ]

[]

[ "en" ]

[ "# Uploaded model\n\n- Developed by: waadarsh\n- License: apache-2.0\n- Finetuned from model : unsloth/llama-3-8b-bnb-4bit\n\nThis llama model was trained 2x faster with Unsloth and Huggingface's TRL library.\n\n<img src=\"URL width=\"200\"/>" ]

[ "TAGS\n#transformers #safetensors #text-generation-inference #unsloth #llama #trl #en #base_model-unsloth/llama-3-8b-bnb-4bit #license-apache-2.0 #endpoints_compatible #region-us \n", "# Uploaded model\n\n- Developed by: waadarsh\n- License: apache-2.0\n- Finetuned from model : unsloth/llama-3-8b-bnb-4bit\n\nThis llama model was trained 2x faster with Unsloth and Huggingface's TRL library.\n\n<img src=\"URL width=\"200\"/>" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2304.01373", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[ "2304.01373", "2101.00027", "2201.07311" ]

[]

[ "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-160M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-160M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-160M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-160M-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-160M-deduped to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-160M-deduped may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-160M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nPythia-160M-deduped was trained on the Pile after the dataset has been globally\ndeduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2304.01373 #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n", "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-160M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-160M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-160M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-160M-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-160M-deduped to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-160M-deduped may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-160M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nPythia-160M-deduped was trained on the Pile after the dataset has been globally\ndeduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "tensorboard", "safetensors", "xlm-roberta", "token-classification", "generated_from_trainer", "base_model:FacebookAI/xlm-roberta-large", "license:mit", "autotrain_compatible", "endpoints_compatible", "region:us" ]

[]

[]

[ "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 1e-05\n* train\\_batch\\_size: 8\n* eval\\_batch\\_size: 8\n* seed: 42\n* gradient\\_accumulation\\_steps: 16\n* total\\_train\\_batch\\_size: 128\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* num\\_epochs: 5", "### Training results", "### Framework versions\n\n\n* Transformers 4.39.0.dev0\n* Pytorch 2.2.1+cu118\n* Datasets 2.18.0\n* Tokenizers 0.15.2" ]

[ "TAGS\n#transformers #tensorboard #safetensors #xlm-roberta #token-classification #generated_from_trainer #base_model-FacebookAI/xlm-roberta-large #license-mit #autotrain_compatible #endpoints_compatible #region-us \n", "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 1e-05\n* train\\_batch\\_size: 8\n* eval\\_batch\\_size: 8\n* seed: 42\n* gradient\\_accumulation\\_steps: 16\n* total\\_train\\_batch\\_size: 128\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* num\\_epochs: 5", "### Training results", "### Framework versions\n\n\n* Transformers 4.39.0.dev0\n* Pytorch 2.2.1+cu118\n* Datasets 2.18.0\n* Tokenizers 0.15.2" ]

[ "transformers", "tensorboard", "safetensors", "distilbert", "text-classification", "generated_from_trainer", "base_model:distilbert-base-uncased", "license:apache-2.0", "autotrain_compatible", "endpoints_compatible", "region:us" ]

[]

[]

[ "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 2e-05\n* train\\_batch\\_size: 16\n* eval\\_batch\\_size: 16\n* seed: 42\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* num\\_epochs: 2", "### Training results", "### Framework versions\n\n\n* Transformers 4.40.0\n* Pytorch 2.2.1+cu121\n* Datasets 2.19.0\n* Tokenizers 0.19.1" ]

[ "TAGS\n#transformers #tensorboard #safetensors #distilbert #text-classification #generated_from_trainer #base_model-distilbert-base-uncased #license-apache-2.0 #autotrain_compatible #endpoints_compatible #region-us \n", "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 2e-05\n* train\\_batch\\_size: 16\n* eval\\_batch\\_size: 16\n* seed: 42\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* num\\_epochs: 2", "### Training results", "### Framework versions\n\n\n* Transformers 4.40.0\n* Pytorch 2.2.1+cu121\n* Datasets 2.19.0\n* Tokenizers 0.19.1" ]

[ "peft", "safetensors", "trl", "sft", "generated_from_trainer", "base_model:meta-llama/Llama-2-7b-chat-hf", "license:llama2", "region:us" ]

[]

[]

[ "# with_board_only_history_with_sys_5epoch_lr1.41e-5\n\nThis model is a fine-tuned version of meta-llama/Llama-2-7b-chat-hf on an unknown dataset.", "## Model description\n\nMore information needed", "## Intended uses & limitations\n\nMore information needed", "## Training and evaluation data\n\nMore information needed", "## Training procedure", "### Training hyperparameters\n\nThe following hyperparameters were used during training:\n- learning_rate: 1.41e-05\n- train_batch_size: 8\n- eval_batch_size: 16\n- seed: 42\n- gradient_accumulation_steps: 16\n- total_train_batch_size: 128\n- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n- lr_scheduler_type: linear\n- num_epochs: 5.0", "### Training results", "### Framework versions\n\n- PEFT 0.10.0\n- Transformers 4.39.2\n- Pytorch 2.2.2\n- Datasets 2.18.0\n- Tokenizers 0.15.2" ]

[ "TAGS\n#peft #safetensors #trl #sft #generated_from_trainer #base_model-meta-llama/Llama-2-7b-chat-hf #license-llama2 #region-us \n", "# with_board_only_history_with_sys_5epoch_lr1.41e-5\n\nThis model is a fine-tuned version of meta-llama/Llama-2-7b-chat-hf on an unknown dataset.", "## Model description\n\nMore information needed", "## Intended uses & limitations\n\nMore information needed", "## Training and evaluation data\n\nMore information needed", "## Training procedure", "### Training hyperparameters\n\nThe following hyperparameters were used during training:\n- learning_rate: 1.41e-05\n- train_batch_size: 8\n- eval_batch_size: 16\n- seed: 42\n- gradient_accumulation_steps: 16\n- total_train_batch_size: 128\n- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n- lr_scheduler_type: linear\n- num_epochs: 5.0", "### Training results", "### Framework versions\n\n- PEFT 0.10.0\n- Transformers 4.39.2\n- Pytorch 2.2.2\n- Datasets 2.18.0\n- Tokenizers 0.15.2" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[ "2101.00027", "2201.07311" ]

[]

[ "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-2.8B-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-2.8B-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-2.8B-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-2.8B-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-2.8B-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-2.8B-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-2.8B-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-2.8B-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n", "### Model Details\n\n\n* Developed by: EleutherAI\n* Model type: Transformer-based Language Model\n* Language: English\n* Learn more: Pythia's GitHub repository\nfor training procedure, config files, and details on how to use.\n* Library: GPT-NeoX\n* License: Apache 2.0\n* Contact: to ask questions about this model, join the EleutherAI\nDiscord, and post them in '#release-discussion'.\nPlease read the existing *Pythia* documentation before asking about it in the\nEleutherAI Discord. For general correspondence: contact@eleuther.\nai.\n\n\n\n\nEngineering details for the *Pythia Suite*. Deduped and \nnon-deduped models of a given size have the same hyperparameters. “Equivalent” \nmodels have **exactly** the same architecture, and the same number of \nnon-embedding parameters.", "### Uses and Limitations", "#### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. To enable the\nstudy of how language models change in the course of training, we provide\n143 evenly spaced intermediate checkpoints per model. These checkpoints are\nhosted on Hugging Face as branches. Note that branch '143000' corresponds\nexactly to the model checkpoint on the 'main' branch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-2.8B-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-2.8B-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "#### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-2.8B-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-2.8B-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “understand” human instructions.", "#### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token deemed statistically most likely by the\nmodel need not produce the most “accurate” text. Never rely on\nPythia-2.8B-deduped to produce factually accurate output.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-2.8B-deduped may produce socially unacceptable or undesirable text,\n*even if* the prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-2.8B-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.", "### Training", "#### Training data\n\n\nPythia-2.8B-deduped was trained on the Pile after the dataset has been\nglobally deduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "#### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for the equivalent of 143000 steps at a batch size\nof 2,097,152 tokens. Two batch sizes were used: 2M and 4M. Models with a batch\nsize of 4M tokens listed were originally trained for 71500 steps instead, with\ncheckpoints every 500 steps. The checkpoints on Hugging Face are renamed for\nconsistency with all 2M batch models, so 'step1000' is the first checkpoint\nfor 'pythia-1.4b' that was saved (corresponding to step 500 in training), and\n'step1000' is likewise the first 'pythia-6.9b' checkpoint that was saved\n(corresponding to 1000 “actual” steps). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.", "### Evaluations\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge – Challenge Set\n\n\n\nSciQ\n", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2304.01373", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "4-bit", "region:us" ]

[ "2304.01373", "2101.00027", "2201.07311" ]

[]

[ "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-1.4B for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-1.4B as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-1.4B has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-1.4B will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-1.4B to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-1.4B may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-1.4B.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-1.4B.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2304.01373 #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #4-bit #region-us \n", "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-1.4B for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-1.4B as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-1.4B has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-1.4B will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-1.4B to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-1.4B may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-1.4B.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-1.4B.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "diffusers", "tensorboard", "safetensors", "text-to-image", "dreambooth", "diffusers-training", "stable-diffusion", "stable-diffusion-diffusers", "base_model:runwayml/stable-diffusion-v1-5", "license:creativeml-openrail-m", "endpoints_compatible", "diffusers:StableDiffusionPipeline", "region:us" ]

[]

[]

[ "# DreamBooth - Ppororo/paper_model\n\nThis is a dreambooth model derived from runwayml/stable-diffusion-v1-5. The weights were trained on a photo of sks paper using DreamBooth.\nYou can find some example images in the following. \n\n\n\nDreamBooth for the text encoder was enabled: False.", "## Intended uses & limitations", "#### How to use", "#### Limitations and bias\n\n[TODO: provide examples of latent issues and potential remediations]", "## Training details\n\n[TODO: describe the data used to train the model]" ]

[ "TAGS\n#diffusers #tensorboard #safetensors #text-to-image #dreambooth #diffusers-training #stable-diffusion #stable-diffusion-diffusers #base_model-runwayml/stable-diffusion-v1-5 #license-creativeml-openrail-m #endpoints_compatible #diffusers-StableDiffusionPipeline #region-us \n", "# DreamBooth - Ppororo/paper_model\n\nThis is a dreambooth model derived from runwayml/stable-diffusion-v1-5. The weights were trained on a photo of sks paper using DreamBooth.\nYou can find some example images in the following. \n\n\n\nDreamBooth for the text encoder was enabled: False.", "## Intended uses & limitations", "#### How to use", "#### Limitations and bias\n\n[TODO: provide examples of latent issues and potential remediations]", "## Training details\n\n[TODO: describe the data used to train the model]" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2304.01373", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[ "2304.01373", "2101.00027", "2201.07311" ]

[]

[ "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-1.4B for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-1.4B as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-1.4B has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-1.4B will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-1.4B to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-1.4B may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-1.4B.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-1.4B.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2304.01373 #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n", "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-1.4B for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-1.4B as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-1.4B has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-1.4B will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-1.4B to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-1.4B may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-1.4B.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-1.4B.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "pegasus", "text2text-generation", "generated_from_trainer", "base_model:google/pegasus-xsum", "autotrain_compatible", "endpoints_compatible", "region:us" ]

[]

[]

[ "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 5e-05\n* train\\_batch\\_size: 4\n* eval\\_batch\\_size: 8\n* seed: 42\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* lr\\_scheduler\\_warmup\\_steps: 500\n* num\\_epochs: 1\n* mixed\\_precision\\_training: Native AMP", "### Training results", "### Framework versions\n\n\n* Transformers 4.40.0\n* Pytorch 2.1.0+cu121\n* Datasets 2.19.0\n* Tokenizers 0.19.1" ]

[ "TAGS\n#transformers #safetensors #pegasus #text2text-generation #generated_from_trainer #base_model-google/pegasus-xsum #autotrain_compatible #endpoints_compatible #region-us \n", "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 5e-05\n* train\\_batch\\_size: 4\n* eval\\_batch\\_size: 8\n* seed: 42\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* lr\\_scheduler\\_warmup\\_steps: 500\n* num\\_epochs: 1\n* mixed\\_precision\\_training: Native AMP", "### Training results", "### Framework versions\n\n\n* Transformers 4.40.0\n* Pytorch 2.1.0+cu121\n* Datasets 2.19.0\n* Tokenizers 0.19.1" ]

[ "transformers", "safetensors", "llama", "text-generation", "mergekit", "merge", "conversational", "base_model:meta-llama/Meta-Llama-3-8B-Instruct", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "region:us" ]

[]

[]

[ "# Untitled Model (1)\n\nThis is a merge of pre-trained language models created using mergekit.", "## Merge Details", "### Merge Method\n\nThis model was merged using the SLERP merge method.", "### Models Merged\n\nThe following models were included in the merge:\n* meta-llama/Meta-Llama-3-8B-Instruct\n* /workspace/ses", "### Configuration\n\nThe following YAML configuration was used to produce this model:" ]

[ "TAGS\n#transformers #safetensors #llama #text-generation #mergekit #merge #conversational #base_model-meta-llama/Meta-Llama-3-8B-Instruct #autotrain_compatible #endpoints_compatible #text-generation-inference #region-us \n", "# Untitled Model (1)\n\nThis is a merge of pre-trained language models created using mergekit.", "## Merge Details", "### Merge Method\n\nThis model was merged using the SLERP merge method.", "### Models Merged\n\nThe following models were included in the merge:\n* meta-llama/Meta-Llama-3-8B-Instruct\n* /workspace/ses", "### Configuration\n\nThe following YAML configuration was used to produce this model:" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2304.01373", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "4-bit", "region:us" ]

[ "2304.01373", "2101.00027", "2201.07311" ]

[]

[ "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-2.8B-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-2.8B-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-2.8B-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-2.8B-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-2.8B-deduped to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-2.8B-deduped may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-2.8B-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nPythia-2.8B-deduped was trained on the Pile after the dataset has been globally\ndeduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2304.01373 #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #4-bit #region-us \n", "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-2.8B-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-2.8B-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-2.8B-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-2.8B-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-2.8B-deduped to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-2.8B-deduped may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-2.8B-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nPythia-2.8B-deduped was trained on the Pile after the dataset has been globally\ndeduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2304.01373", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "4-bit", "region:us" ]

[ "2304.01373", "2101.00027", "2201.07311" ]

[]

[ "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-70M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-70M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-70M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-70M-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-70M-deduped to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-70M-deduped may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-70M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nPythia-70M-deduped was trained on the Pile after the dataset has been globally\ndeduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2304.01373 #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #4-bit #region-us \n", "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-70M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-70M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-70M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-70M-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-70M-deduped to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-70M-deduped may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-70M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nPythia-70M-deduped was trained on the Pile after the dataset has been globally\ndeduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2304.01373", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[ "2304.01373", "2101.00027", "2201.07311" ]

[]

[ "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-70M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-70M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-70M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-70M-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-70M-deduped to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-70M-deduped may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-70M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nPythia-70M-deduped was trained on the Pile after the dataset has been globally\ndeduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2304.01373 #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n", "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-70M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-70M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-70M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-70M-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-70M-deduped to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-70M-deduped may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-70M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nPythia-70M-deduped was trained on the Pile after the dataset has been globally\ndeduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "trl", "dpo", "arxiv:1910.09700", "endpoints_compatible", "region:us" ]

[ "1910.09700" ]

[]

[ "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a transformers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "TAGS\n#transformers #safetensors #trl #dpo #arxiv-1910.09700 #endpoints_compatible #region-us \n", "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a transformers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2304.01373", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[ "2304.01373", "2101.00027", "2201.07311" ]

[]

[ "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-2.8B-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-2.8B-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-2.8B-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-2.8B-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-2.8B-deduped to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-2.8B-deduped may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-2.8B-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nPythia-2.8B-deduped was trained on the Pile after the dataset has been globally\ndeduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2304.01373 #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n", "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-2.8B-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-2.8B-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-2.8B-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-2.8B-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-2.8B-deduped to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-2.8B-deduped may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-2.8B-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nPythia-2.8B-deduped was trained on the Pile after the dataset has been globally\ndeduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "vision-encoder-decoder", "validation_1", "validation_2", "validation_3", "validation_4", "validation_5", "validation_6", "validation_7", "arxiv:1910.09700", "endpoints_compatible", "region:us" ]

[ "1910.09700" ]

[]

[ "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a transformers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "TAGS\n#transformers #safetensors #vision-encoder-decoder #validation_1 #validation_2 #validation_3 #validation_4 #validation_5 #validation_6 #validation_7 #arxiv-1910.09700 #endpoints_compatible #region-us \n", "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a transformers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "transformers", "safetensors", "arxiv:1910.09700", "endpoints_compatible", "region:us" ]

[ "1910.09700" ]

[]

[ "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a transformers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "TAGS\n#transformers #safetensors #arxiv-1910.09700 #endpoints_compatible #region-us \n", "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a transformers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2304.01373", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "4-bit", "region:us" ]

[ "2304.01373", "2101.00027", "2201.07311" ]

[]

[ "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-2.8B for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-2.8B as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-2.8B has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-2.8B will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-2.8B to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-2.8B may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-2.8B.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-2.8B.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2304.01373 #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #4-bit #region-us \n", "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-2.8B for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-2.8B as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-2.8B has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-2.8B will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-2.8B to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-2.8B may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-2.8B.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-2.8B.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2304.01373", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "4-bit", "region:us" ]

[ "2304.01373", "2101.00027", "2201.07311" ]

[]

[ "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-410M for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-410M as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-410M has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-410M will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-410M to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-410M may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-410M.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-410M.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2304.01373 #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #4-bit #region-us \n", "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-410M for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-410M as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-410M has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-410M will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-410M to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-410M may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-410M.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-410M.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "setfit", "safetensors", "bert", "sentence-transformers", "text-classification", "generated_from_setfit_trainer", "arxiv:2209.11055", "base_model:BAAI/bge-small-en-v1.5", "model-index", "region:us" ]

[ "2209.11055" ]

[]

[ "### Model Description\n\n\n* Model Type: SetFit\n* Sentence Transformer body: BAAI/bge-small-en-v1.5\n* Classification head: a LogisticRegression instance\n* Maximum Sequence Length: 512 tokens\n* Number of Classes: 5 classes", "### Model Sources\n\n\n* Repository: SetFit on GitHub\n* Paper: Efficient Few-Shot Learning Without Prompts\n* Blogpost: SetFit: Efficient Few-Shot Learning Without Prompts", "### Model Labels\n\n\n\nEvaluation\n----------", "### Metrics\n\n\n\nUses\n----", "### Direct Use for Inference\n\n\nFirst install the SetFit library:\n\n\nThen you can load this model and run inference.\n\n\nTraining Details\n----------------", "### Training Set Metrics", "### Training Hyperparameters\n\n\n* batch\\_size: (32, 32)\n* num\\_epochs: (10, 10)\n* max\\_steps: -1\n* sampling\\_strategy: oversampling\n* body\\_learning\\_rate: (2e-05, 1e-05)\n* head\\_learning\\_rate: 0.01\n* loss: CosineSimilarityLoss\n* distance\\_metric: cosine\\_distance\n* margin: 0.25\n* end\\_to\\_end: False\n* use\\_amp: False\n* warmup\\_proportion: 0.1\n* seed: 42\n* eval\\_max\\_steps: -1\n* load\\_best\\_model\\_at\\_end: False", "### Training Results", "### Framework Versions\n\n\n* Python: 3.11.8\n* SetFit: 1.0.3\n* Sentence Transformers: 2.7.0\n* Transformers: 4.40.0\n* PyTorch: 2.2.2\n* Datasets: 2.19.0\n* Tokenizers: 0.19.1", "### BibTeX" ]

[ "TAGS\n#setfit #safetensors #bert #sentence-transformers #text-classification #generated_from_setfit_trainer #arxiv-2209.11055 #base_model-BAAI/bge-small-en-v1.5 #model-index #region-us \n", "### Model Description\n\n\n* Model Type: SetFit\n* Sentence Transformer body: BAAI/bge-small-en-v1.5\n* Classification head: a LogisticRegression instance\n* Maximum Sequence Length: 512 tokens\n* Number of Classes: 5 classes", "### Model Sources\n\n\n* Repository: SetFit on GitHub\n* Paper: Efficient Few-Shot Learning Without Prompts\n* Blogpost: SetFit: Efficient Few-Shot Learning Without Prompts", "### Model Labels\n\n\n\nEvaluation\n----------", "### Metrics\n\n\n\nUses\n----", "### Direct Use for Inference\n\n\nFirst install the SetFit library:\n\n\nThen you can load this model and run inference.\n\n\nTraining Details\n----------------", "### Training Set Metrics", "### Training Hyperparameters\n\n\n* batch\\_size: (32, 32)\n* num\\_epochs: (10, 10)\n* max\\_steps: -1\n* sampling\\_strategy: oversampling\n* body\\_learning\\_rate: (2e-05, 1e-05)\n* head\\_learning\\_rate: 0.01\n* loss: CosineSimilarityLoss\n* distance\\_metric: cosine\\_distance\n* margin: 0.25\n* end\\_to\\_end: False\n* use\\_amp: False\n* warmup\\_proportion: 0.1\n* seed: 42\n* eval\\_max\\_steps: -1\n* load\\_best\\_model\\_at\\_end: False", "### Training Results", "### Framework Versions\n\n\n* Python: 3.11.8\n* SetFit: 1.0.3\n* Sentence Transformers: 2.7.0\n* Transformers: 4.40.0\n* PyTorch: 2.2.2\n* Datasets: 2.19.0\n* Tokenizers: 0.19.1", "### BibTeX" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2304.01373", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[ "2304.01373", "2101.00027", "2201.07311" ]

[]

[ "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-410M for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-410M as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-410M has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-410M will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-410M to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-410M may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-410M.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-410M.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2304.01373 #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n", "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-410M for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-410M as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-410M has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-410M will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-410M to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-410M may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-410M.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-410M.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2304.01373", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "4-bit", "region:us" ]

[ "2304.01373", "2101.00027", "2201.07311" ]

[]

[ "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-410M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-410M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-410M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means XNPythia-410M-dedupedAME will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-410M-deduped to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-410M-deduped may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-410M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nPythia-410M-deduped was trained on the Pile after the dataset has been globally\ndeduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2304.01373 #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #4-bit #region-us \n", "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-410M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-410M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-410M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means XNPythia-410M-dedupedAME will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-410M-deduped to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-410M-deduped may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-410M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nPythia-410M-deduped was trained on the Pile after the dataset has been globally\ndeduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "trl", "dpo", "arxiv:1910.09700", "endpoints_compatible", "region:us" ]

[ "1910.09700" ]

[]

[ "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a transformers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "TAGS\n#transformers #safetensors #trl #dpo #arxiv-1910.09700 #endpoints_compatible #region-us \n", "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a transformers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "transformers", "safetensors", "llama", "text-generation", "mergekit", "merge", "conversational", "base_model:abacusai/Llama-3-Smaug-8B", "base_model:cognitivecomputations/dolphin-2.9-llama3-8b", "license:llama3", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "region:us" ]

[]

[]

[ "# Llama-3-Smolphin-8b\n\n<figure>\n\n!Smolphin \n\n</figure>\n\nThis is a merge of pre-trained language models URL and URL created using mergekit.", "## Merge Details", "### Merge Method\n\nThis model was merged using the SLERP merge method.", "### Models Merged\n\nThe following models were included in the merge:\n* URL\n* URL", "### Configuration\n\nThe following YAML configuration was used to produce this model:" ]

[ "TAGS\n#transformers #safetensors #llama #text-generation #mergekit #merge #conversational #base_model-abacusai/Llama-3-Smaug-8B #base_model-cognitivecomputations/dolphin-2.9-llama3-8b #license-llama3 #autotrain_compatible #endpoints_compatible #text-generation-inference #region-us \n", "# Llama-3-Smolphin-8b\n\n<figure>\n\n!Smolphin \n\n</figure>\n\nThis is a merge of pre-trained language models URL and URL created using mergekit.", "## Merge Details", "### Merge Method\n\nThis model was merged using the SLERP merge method.", "### Models Merged\n\nThe following models were included in the merge:\n* URL\n* URL", "### Configuration\n\nThe following YAML configuration was used to produce this model:" ]

[ "peft", "safetensors", "codegen", "arxiv:1910.09700", "base_model:Salesforce/codegen-350M-mono", "region:us" ]

[ "1910.09700" ]

[]

[ "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\n\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact", "### Framework versions\n\n- PEFT 0.10.0" ]

[ "TAGS\n#peft #safetensors #codegen #arxiv-1910.09700 #base_model-Salesforce/codegen-350M-mono #region-us \n", "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\n\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact", "### Framework versions\n\n- PEFT 0.10.0" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2304.01373", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[ "2304.01373", "2101.00027", "2201.07311" ]

[]

[ "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-410M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-410M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-410M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means XNPythia-410M-dedupedAME will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-410M-deduped to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-410M-deduped may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-410M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nPythia-410M-deduped was trained on the Pile after the dataset has been globally\ndeduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2304.01373 #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n", "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-410M-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-410M-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-410M-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means XNPythia-410M-dedupedAME will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-410M-deduped to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-410M-deduped may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-410M-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nPythia-410M-deduped was trained on the Pile after the dataset has been globally\ndeduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "diffusers", "safetensors", "arxiv:1910.09700", "region:us" ]

[ "1910.09700" ]

[]

[ "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a diffusers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "TAGS\n#diffusers #safetensors #arxiv-1910.09700 #region-us \n", "# Model Card for Model ID", "## Model Details", "### Model Description\n\n\n\nThis is the model card of a diffusers model that has been pushed on the Hub. This model card has been automatically generated.\n\n- Developed by: \n- Funded by [optional]: \n- Shared by [optional]: \n- Model type: \n- Language(s) (NLP): \n- License: \n- Finetuned from model [optional]:", "### Model Sources [optional]\n\n\n\n- Repository: \n- Paper [optional]: \n- Demo [optional]:", "## Uses", "### Direct Use", "### Downstream Use [optional]", "### Out-of-Scope Use", "## Bias, Risks, and Limitations", "### Recommendations\n\n\n\nUsers (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.", "## How to Get Started with the Model\n\nUse the code below to get started with the model.", "## Training Details", "### Training Data", "### Training Procedure", "#### Preprocessing [optional]", "#### Training Hyperparameters\n\n- Training regime:", "#### Speeds, Sizes, Times [optional]", "## Evaluation", "### Testing Data, Factors & Metrics", "#### Testing Data", "#### Factors", "#### Metrics", "### Results", "#### Summary", "## Model Examination [optional]", "## Environmental Impact\n\n\n\nCarbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).\n\n- Hardware Type: \n- Hours used: \n- Cloud Provider: \n- Compute Region: \n- Carbon Emitted:", "## Technical Specifications [optional]", "### Model Architecture and Objective", "### Compute Infrastructure", "#### Hardware", "#### Software\n\n\n\n[optional]\n\n\n\nBibTeX:\n\n\n\nAPA:", "## Glossary [optional]", "## More Information [optional]", "## Model Card Authors [optional]", "## Model Card Contact" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2304.01373", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[ "2304.01373", "2101.00027", "2201.07311" ]

[]

[ "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-2.8B for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-2.8B as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-2.8B has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-2.8B will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-2.8B to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-2.8B may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-2.8B.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-2.8B.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2304.01373 #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n", "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-2.8B for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-2.8B as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-2.8B has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-2.8B will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-2.8B to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-2.8B may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-2.8B.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror. \n\nThe Pile was not deduplicated before being used to train Pythia-2.8B.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "pytorch", "safetensors", "megatron-bert", "feature-extraction", "sv", "endpoints_compatible", "region:us" ]

[]

[ "sv" ]

[ "# URL-32k-no_pretok.25k-steps\n\nThis BERT model was trained using the NeMo library.\nThe size of the model is a regular bert-large.\nThe model was trained on more than 245GB of data, consisting mostly of web-data and Swedish newspaper text curated by the National Library of Sweden.\n\nTraining was done for 25k training steps using a batch size of 8k.\n\nThe model has multiple sibling models trained on the same dataset using different tokenizers or more/less parameters:\n- URL-32k-no_pretok.25k-steps\n- URL-64k-no_pretok.25k-steps\n- URL-bpe-32k-no_pretok.25k-steps\n- URL-bpe-32k-pretok.25k-steps\n- URL-bpe-64k-no_pretok.25k-steps\n- URL-bpe-64k-pretok.25k-steps\n- URL-base.unigram-32k-no_pretok.25k-steps\n- URL-base.unigram-32k-pretok.25k-steps\n- URL-base.unigram-64k-no_pretok.25k-steps\n- URL-base.unigram-64k-pretok.25k-steps\n- URL-base.wordpiece-32k-no_pretok.25k-steps\n- URL-base.wordpiece-32k-pretok.25k-steps\n- URL-base.wordpiece-64k-no_pretok.25k-steps\n- URL-base.wordpiece-64k-pretok.25k-steps\n- URL-64k-no_pretok.25k-steps\n- URL-bpe-32k-pretok.25k-steps\n- URL-large.unigram-32k-pretok.25k-steps\n- URL-large.unigram-64k-pretok.25k-steps\n- URL-large.wordpiece-32k-pretok.25k-steps\n- URL-large.wordpiece-64k-pretok.25k-steps", "## Acknowledgements\n\nThe training was performed on the Luxembourg national supercomputer MeluXina.\nThe authors gratefully acknowledge the LuxProvide teams for their expert support." ]

[ "TAGS\n#transformers #pytorch #safetensors #megatron-bert #feature-extraction #sv #endpoints_compatible #region-us \n", "# URL-32k-no_pretok.25k-steps\n\nThis BERT model was trained using the NeMo library.\nThe size of the model is a regular bert-large.\nThe model was trained on more than 245GB of data, consisting mostly of web-data and Swedish newspaper text curated by the National Library of Sweden.\n\nTraining was done for 25k training steps using a batch size of 8k.\n\nThe model has multiple sibling models trained on the same dataset using different tokenizers or more/less parameters:\n- URL-32k-no_pretok.25k-steps\n- URL-64k-no_pretok.25k-steps\n- URL-bpe-32k-no_pretok.25k-steps\n- URL-bpe-32k-pretok.25k-steps\n- URL-bpe-64k-no_pretok.25k-steps\n- URL-bpe-64k-pretok.25k-steps\n- URL-base.unigram-32k-no_pretok.25k-steps\n- URL-base.unigram-32k-pretok.25k-steps\n- URL-base.unigram-64k-no_pretok.25k-steps\n- URL-base.unigram-64k-pretok.25k-steps\n- URL-base.wordpiece-32k-no_pretok.25k-steps\n- URL-base.wordpiece-32k-pretok.25k-steps\n- URL-base.wordpiece-64k-no_pretok.25k-steps\n- URL-base.wordpiece-64k-pretok.25k-steps\n- URL-64k-no_pretok.25k-steps\n- URL-bpe-32k-pretok.25k-steps\n- URL-large.unigram-32k-pretok.25k-steps\n- URL-large.unigram-64k-pretok.25k-steps\n- URL-large.wordpiece-32k-pretok.25k-steps\n- URL-large.wordpiece-64k-pretok.25k-steps", "## Acknowledgements\n\nThe training was performed on the Luxembourg national supercomputer MeluXina.\nThe authors gratefully acknowledge the LuxProvide teams for their expert support." ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2304.01373", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "4-bit", "region:us" ]

[ "2304.01373", "2101.00027", "2201.07311" ]

[]

[ "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-1B-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-1B-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-1B-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-1B-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-1B-deduped to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-1B-deduped may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-1B-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nPythia-1B-deduped was trained on the Pile after the dataset has been globally\ndeduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2304.01373 #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #4-bit #region-us \n", "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-1B-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-1B-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-1B-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-1B-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-1B-deduped to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-1B-deduped may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-1B-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nPythia-1B-deduped was trained on the Pile after the dataset has been globally\ndeduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "4-bit", "region:us" ]

[]

[]

[]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #autotrain_compatible #endpoints_compatible #text-generation-inference #4-bit #region-us \n" ]

[ "xtuner", "gguf", "llama-cpp", "gguf-my-repo", "image-text-to-text", "dataset:Lin-Chen/ShareGPT4V", "region:us" ]

[]

[]

[ "# stormchaser/llava-llama-3-8b-v1_1-Q6_K-GGUF\nThis model was converted to GGUF format from 'xtuner/llava-llama-3-8b-v1_1' using URL via the URL's GGUF-my-repo space.\nRefer to the original model card for more details on the model.", "## Use with URL\n\nInstall URL through brew.\n\n\nInvoke the URL server or the CLI.\n\nCLI:\n\n\n\nServer:\n\n\n\nNote: You can also use this checkpoint directly through the usage steps listed in the URL repo as well." ]

[ "TAGS\n#xtuner #gguf #llama-cpp #gguf-my-repo #image-text-to-text #dataset-Lin-Chen/ShareGPT4V #region-us \n", "# stormchaser/llava-llama-3-8b-v1_1-Q6_K-GGUF\nThis model was converted to GGUF format from 'xtuner/llava-llama-3-8b-v1_1' using URL via the URL's GGUF-my-repo space.\nRefer to the original model card for more details on the model.", "## Use with URL\n\nInstall URL through brew.\n\n\nInvoke the URL server or the CLI.\n\nCLI:\n\n\n\nServer:\n\n\n\nNote: You can also use this checkpoint directly through the usage steps listed in the URL repo as well." ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[]

[]

[]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n" ]

[ "transformers", "tensorboard", "safetensors", "albert", "text-classification", "generated_from_trainer", "base_model:KalaiselvanD/kalai_bert_model_test_3_out", "license:apache-2.0", "autotrain_compatible", "endpoints_compatible", "region:us" ]

[]

[]

[ "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 2e-05\n* train\\_batch\\_size: 16\n* eval\\_batch\\_size: 16\n* seed: 42\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* num\\_epochs: 3", "### Training results", "### Framework versions\n\n\n* Transformers 4.40.0\n* Pytorch 2.2.1+cu121\n* Datasets 2.19.0\n* Tokenizers 0.19.1" ]

[ "TAGS\n#transformers #tensorboard #safetensors #albert #text-classification #generated_from_trainer #base_model-KalaiselvanD/kalai_bert_model_test_3_out #license-apache-2.0 #autotrain_compatible #endpoints_compatible #region-us \n", "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 2e-05\n* train\\_batch\\_size: 16\n* eval\\_batch\\_size: 16\n* seed: 42\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* num\\_epochs: 3", "### Training results", "### Framework versions\n\n\n* Transformers 4.40.0\n* Pytorch 2.2.1+cu121\n* Datasets 2.19.0\n* Tokenizers 0.19.1" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "arxiv:2304.01373", "arxiv:2101.00027", "arxiv:2201.07311", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "8-bit", "region:us" ]

[ "2304.01373", "2101.00027", "2201.07311" ]

[]

[ "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-1B-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-1B-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-1B-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-1B-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-1B-deduped to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-1B-deduped may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-1B-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nPythia-1B-deduped was trained on the Pile after the dataset has been globally\ndeduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #arxiv-2304.01373 #arxiv-2101.00027 #arxiv-2201.07311 #autotrain_compatible #endpoints_compatible #text-generation-inference #8-bit #region-us \n", "### Intended Use\n\n\nThe primary intended use of Pythia is research on the behavior, functionality,\nand limitations of large language models. This suite is intended to provide\na controlled setting for performing scientific experiments. We also provide\n154 checkpoints per model: initial 'step0', 10 log-spaced checkpoints\n'step{1,2,4...512}', and 143 evenly-spaced checkpoints from 'step1000' to\n'step143000'. These checkpoints are hosted on Hugging Face as branches. Note\nthat branch '143000' corresponds exactly to the model checkpoint on the 'main'\nbranch of each model.\n\n\nYou may also further fine-tune and adapt Pythia-1B-deduped for deployment,\nas long as your use is in accordance with the Apache 2.0 license. Pythia\nmodels work with the Hugging Face Transformers\nLibrary. If you decide to use\npre-trained Pythia-1B-deduped as a basis for your fine-tuned model, please\nconduct your own risk and bias assessment.", "### Out-of-scope use\n\n\nThe Pythia Suite is not intended for deployment. It is not a in itself\na product and cannot be used for human-facing interactions. For example,\nthe model may generate harmful or offensive text. Please evaluate the risks\nassociated with your particular use case.\n\n\nPythia models are English-language only, and are not suitable for translation\nor generating text in other languages.\n\n\nPythia-1B-deduped has not been fine-tuned for downstream contexts in which\nlanguage models are commonly deployed, such as writing genre prose,\nor commercial chatbots. This means Pythia-1B-deduped will not\nrespond to a given prompt the way a product like ChatGPT does. This is because,\nunlike this model, ChatGPT was fine-tuned using methods such as Reinforcement\nLearning from Human Feedback (RLHF) to better “follow” human instructions.", "### Limitations and biases\n\n\nThe core functionality of a large language model is to take a string of text\nand predict the next token. The token used by the model need not produce the\nmost “accurate” text. Never rely on Pythia-1B-deduped to produce factually accurate\noutput.\n\n\nThis model was trained on the Pile, a dataset\nknown to contain profanity and texts that are lewd or otherwise offensive.\nSee Section 6 of the Pile paper for a\ndiscussion of documented biases with regards to gender, religion, and race.\nPythia-1B-deduped may produce socially unacceptable or undesirable text, *even if*\nthe prompt itself does not include anything explicitly offensive.\n\n\nIf you plan on using text generated through, for example, the Hosted Inference\nAPI, we recommend having a human curate the outputs of this language model\nbefore presenting it to other people. Please inform your audience that the\ntext was generated by Pythia-1B-deduped.", "### Quickstart\n\n\nPythia models can be loaded and used via the following code, demonstrated here\nfor the third 'pythia-70m-deduped' checkpoint:\n\n\nRevision/branch 'step143000' corresponds exactly to the model checkpoint on\nthe 'main' branch of each model. \n\nFor more information on how to use all Pythia models, see documentation on\nGitHub.\n\n\nTraining\n--------", "### Training data\n\n\nPythia-1B-deduped was trained on the Pile after the dataset has been globally\ndeduplicated. \n\nThe Pile is a 825GiB general-purpose dataset in\nEnglish. It was created by EleutherAI specifically for training large language\nmodels. It contains texts from 22 diverse sources, roughly broken down into\nfive categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl),\nprose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and\nmiscellaneous (e.g. GitHub, Enron Emails). See the Pile\npaper for a breakdown of all data sources,\nmethodology, and a discussion of ethical implications. Consult the\ndatasheet for more detailed documentation\nabout the Pile and its component datasets. The Pile can be downloaded from\nthe official website, or from a community\nmirror.", "### Training procedure\n\n\nAll models were trained on the exact same data, in the exact same order. Each\nmodel saw 299,892,736,000 tokens during training, and 143 checkpoints for each\nmodel are saved every 2,097,152,000 tokens, spaced evenly throughout training,\nfrom 'step1000' to 'step143000' (which is the same as 'main'). In addition, we\nalso provide frequent early checkpoints: 'step0' and 'step{1,2,4...512}'.\nThis corresponds to training for just under 1 epoch on the Pile for\nnon-deduplicated models, and about 1.5 epochs on the deduplicated Pile.\n\n\nAll *Pythia* models trained for 143000 steps at a batch size\nof 2M (2,097,152 tokens). \n\nSee GitHub for more details on training\nprocedure, including how to reproduce\nit. \n\nPythia uses the same tokenizer as GPT-NeoX-\n20B.\n\n\nEvaluations\n-----------\n\n\nAll 16 *Pythia* models were evaluated using the LM Evaluation\nHarness. You can access\nthe results by model and step at 'results/json/\\*' in the GitHub\nrepository. \n\nExpand the sections below to see plots of evaluation results for all\nPythia and Pythia-deduped models compared with OPT and BLOOM.\n\n\n\nLAMBADA – OpenAI\n\n\n\nPhysical Interaction: Question Answering (PIQA)\n\n\n\nWinoGrande\n\n\n\nAI2 Reasoning Challenge—Easy Set\n\n\n\nSciQ\n\n\nChangelog\n---------\n\n\nThis section compares differences between previously released\nPythia v0 and the current\nmodels. See Appendix B of the Pythia paper for further discussion of these\nchanges and the motivation behind them. We found that retraining Pythia had no\nimpact on benchmark performance.\n\n\n* All model sizes are now trained with uniform batch size of 2M tokens.\nPreviously, the models of size 160M, 410M, and 1.4B parameters were trained\nwith batch sizes of 4M tokens.\n* We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64,\n128,256,512} in addition to every 1000 training steps.\n* Flash Attention was used in the new retrained suite.\n* We remedied a minor inconsistency that existed in the original suite: all\nmodels of size 2.8B parameters or smaller had a learning rate (LR) schedule\nwhich decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and\n12B models all used an LR schedule which decayed to a minimum LR of 0. In\nthe redone training runs, we rectified this inconsistency: all models now were\ntrained with LR decaying to a minimum of 0.1× their maximum LR.", "### Naming convention and parameter count\n\n\n*Pythia* models were renamed in January 2023. It is possible that the old\nnaming convention still persists in some documentation by accident. The\ncurrent naming convention (70M, 160M, etc.) is based on total parameter count." ]

[ "peft", "safetensors", "generated_from_trainer", "base_model:TheBloke/Llama-2-7B-fp16", "region:us" ]

[]

[]

[ "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 0.0003\n* train\\_batch\\_size: 2\n* eval\\_batch\\_size: 8\n* seed: 42\n* gradient\\_accumulation\\_steps: 10\n* total\\_train\\_batch\\_size: 20\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* training\\_steps: 30", "### Training results", "### Framework versions\n\n\n* PEFT 0.10.0\n* Transformers 4.39.3\n* Pytorch 2.2.2+cu121\n* Datasets 2.18.0\n* Tokenizers 0.15.2" ]

[ "TAGS\n#peft #safetensors #generated_from_trainer #base_model-TheBloke/Llama-2-7B-fp16 #region-us \n", "### Training hyperparameters\n\n\nThe following hyperparameters were used during training:\n\n\n* learning\\_rate: 0.0003\n* train\\_batch\\_size: 2\n* eval\\_batch\\_size: 8\n* seed: 42\n* gradient\\_accumulation\\_steps: 10\n* total\\_train\\_batch\\_size: 20\n* optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08\n* lr\\_scheduler\\_type: linear\n* training\\_steps: 30", "### Training results", "### Framework versions\n\n\n* PEFT 0.10.0\n* Transformers 4.39.3\n* Pytorch 2.2.2+cu121\n* Datasets 2.18.0\n* Tokenizers 0.15.2" ]

[ "transformers", "safetensors", "gpt_neox", "text-generation", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "4-bit", "region:us" ]

[]

[]

[]

[ "TAGS\n#transformers #safetensors #gpt_neox #text-generation #autotrain_compatible #endpoints_compatible #text-generation-inference #4-bit #region-us \n" ]