Advancing Open-source Language Models with Mixed-Quality Data

OPENCHAT3.5 1210
🏆 The Overall Best Performing Open Source 7B Model 🏆
🤖 Outperforms ChatGPT (March) and Grok-1 🤖
🚀15-point improvement in Coding over OpenChat-3.5🚀

New Features
💡 2 Modes: Coding + Generalist, Mathematical Reasoning 💡
🧑‍⚖️ Experimental support for Evaluator and Feedback capabilities 🧑‍⚖️

Usage
Benchmarks
Limitations
License
Dataset Details
Citation
Acknowledgements

Usage

To use this model, we highly recommend installing the OpenChat package by following the installation guide in our repository and using the OpenChat OpenAI-compatible API server by running the serving command from the table below. The server is optimized for high-throughput deployment using vLLM and can run on a consumer GPU with 24GB RAM. To enable tensor parallelism, append --tensor-parallel-size N to the serving command.

Once started, the server listens at localhost:18888 for requests and is compatible with the OpenAI ChatCompletion API specifications. Please refer to the example request below for reference. Additionally, you can use the OpenChat Web UI for a user-friendly experience.

If you want to deploy the server as an online service, you can use --api-keys sk-KEY1 sk-KEY2 ... to specify allowed API keys and --disable-log-requests --disable-log-stats --log-file openchat.log for logging only to a file. For security purposes, we recommend using an HTTPS gateway in front of the server.

Model	Size	Context	Weights	Serving
OpenChat 3.5 1210	7B	8192	Huggingface	`python -m ochat.serving.openai_api_server --model openchat/openchat_3.5_1210 --engine-use-ray --worker-use-ray`

Example request (click to expand)

💡 Default Mode (GPT4 Correct): Best for coding, chat and general tasks

curl http://localhost:18888/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openchat_3.5",
    "messages": [{"role": "user", "content": "You are a large language model named OpenChat. Write a poem to describe yourself"}]
  }'

🧮 Mathematical Reasoning Mode: Tailored for solving math problems

curl http://localhost:18888/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openchat_3.5",
    "condition": "Math Correct",
    "messages": [{"role": "user", "content": "10.3 − 7988.8133 = "}]
  }'

Conversation templates

💡 Default Mode (GPT4 Correct): Best for coding, chat and general tasks

GPT4 Correct User: Hello<|end_of_turn|>GPT4 Correct Assistant: Hi<|end_of_turn|>GPT4 Correct User: How are you today?<|end_of_turn|>GPT4 Correct Assistant:

🧮 Mathematical Reasoning Mode: Tailored for solving math problems

Math Correct User: 10.3 − 7988.8133=<|end_of_turn|>Math Correct Assistant:

⚠️ Notice: Remember to set <|end_of_turn|> as end of generation token.

The default (GPT4 Correct) template is also available as the integrated tokenizer.chat_template, which can be used instead of manually specifying the template:

messages = [
    {"role": "user", "content": "Hello"},
    {"role": "assistant", "content": "Hi"},
    {"role": "user", "content": "How are you today?"}
]
tokens = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
assert tokens == [1, 420, 6316, 28781, 3198, 3123, 1247, 28747, 22557, 32000, 420, 6316, 28781, 3198, 3123, 21631, 28747, 15359, 32000, 420, 6316, 28781, 3198, 3123, 1247, 28747, 1602, 460, 368, 3154, 28804, 32000, 420, 6316, 28781, 3198, 3123, 21631, 28747]

(Experimental) Evaluator / Feedback Capabilities

We've included evaluator capabilities in this release to advance open-source models as evaluators. You can use `Default Mode (GPT4 Correct)` with the following prompt (same as [Prometheus](https://huggingface.co/datasets/kaist-ai/Feedback-Collection)) to evaluate a response.

###Task Description:
An instruction (might include an Input inside it), a response to evaluate, a reference answer that gets a score of 5, and a score rubric representing a evaluation criteria are given.
1. Write a detailed feedback that assess the quality of the response strictly based on the given score rubric, not evaluating in general.
2. After writing a feedback, write a score that is an integer between 1 and 5. You should refer to the score rubric.
3. The output format should look as follows: "Feedback: (write a feedback for criteria) [RESULT] (an integer number between 1 and 5)"
4. Please do not generate any other opening, closing, and explanations.

###The instruction to evaluate:
{orig_instruction}

###Response to evaluate:
{orig_response}

###Reference Answer (Score 5):
{orig_reference_answer}

###Score Rubrics:
[{orig_criteria}]
Score 1: {orig_score1_description}
Score 2: {orig_score2_description}
Score 3: {orig_score3_description}
Score 4: {orig_score4_description}
Score 5: {orig_score5_description}

###Feedback:

Benchmarks

Model	# Params	Average	MT-Bench	HumanEval	BBH MC	AGIEval	TruthfulQA	MMLU	GSM8K	BBH CoT
OpenChat-3.5-1210	7B	63.8	7.76	68.9	49.5	48.0	61.8	65.3	77.3	61.8
OpenChat-3.5	7B	61.6	7.81	55.5	47.6	47.4	59.1	64.3	77.3	63.5
ChatGPT (March)*	?	61.5	7.94	48.1	47.6	47.1	57.7	67.3	74.9	70.1

OpenHermes 2.5	7B	59.3	7.54	48.2	49.4	46.5	57.5	63.8	73.5	59.9
OpenOrca Mistral	7B	52.7	6.86	38.4	49.4	42.9	45.9	59.3	59.1	58.1
Zephyr-β^	7B	34.6	7.34	22.0	40.6	39.0	40.8	39.8	5.1	16.0
Mistral	7B	-	6.84	30.5	39.0	38.0	-	60.1	52.2	-

Evaluation Details(click to expand)

*: ChatGPT (March) results are from [GPT-4 Technical Report](https://arxiv.org/abs/2303.08774), [Chain-of-Thought Hub](https://github.com/FranxYao/chain-of-thought-hub), and our evaluation. Please note that ChatGPT is not a fixed baseline and evolves rapidly over time.

^: Zephyr-β often fails to follow few-shot CoT instructions, likely because it was aligned with only chat data but not trained on few-shot data.

**: Mistral and Open-source SOTA results are taken from reported results in instruction-tuned model papers and official repositories.

All models are evaluated in chat mode (e.g. with the respective conversation template applied). All zero-shot benchmarks follow the same setting as in the AGIEval paper and Orca paper. CoT tasks use the same configuration as Chain-of-Thought Hub, HumanEval is evaluated with EvalPlus, and MT-bench is run using FastChat. To reproduce our results, follow the instructions in our repository.

HumanEval+

Model	Size	HumanEval+ pass@1
ChatGPT (December 12, 2023)	-	64.6
WizardCoder-Python-34B-V1.0	34B	64.6
OpenChat 3.5 (Dec 10)	7B	63.4
OpenHermes 2.5	7B	41.5

OpenChat-3.5-1210 vs. Grok

	License	# Param	Average	MMLU	HumanEval	MATH	GSM8k
OpenChat 3.5 1210	Apache-2.0	7B	60.1	65.3	68.9	28.9	77.3
OpenChat 3.5	Apache-2.0	7B	56.4	64.3	55.5	28.6	77.3
Grok-0	Proprietary	33B	44.5	65.7	39.7	15.7	56.8
Grok-1	Proprietary	???B	55.8	73	63.2	23.9	62.9

*: Grok results are reported by X.AI.

中文评估结果 / Chinese Evaluations

⚠️ Note that this model was not explicitly trained in Chinese (only < 0.1% of the data is in Chinese). 请注意本模型没有针对性训练中文（中文数据占比小于0.1%）。

Multi-Level Multi-Discipline Chinese Evaluation Suite (CEVAL)

Model	Avg	STEM	Social Science	Humanities	Others
ChatGPT	54.4	52.9	61.8	50.9	53.6
OpenChat	47.29	45.22	52.49	48.52	45.08

Massive Multitask Language Understanding in Chinese (CMMLU, 5-shot)

Models	STEM	Humanities	SocialSciences	Other	ChinaSpecific	Avg
ChatGPT	47.81	55.68	56.5	62.66	50.69	55.51
OpenChat	38.7	45.99	48.32	50.23	43.27	45.85

Limitations

Foundation Model Limitations Despite its advanced capabilities, OpenChat is still bound by the limitations inherent in its foundation models. These limitations may impact the model's performance in areas such as:

Complex reasoning
Mathematical and arithmetic tasks
Programming and coding challenges

Hallucination of Non-existent Information OpenChat may sometimes generate information that does not exist or is not accurate, also known as "hallucination". Users should be aware of this possibility and verify any critical information obtained from the model.

Safety OpenChat may sometimes generate harmful, hate speech, biased responses, or answer unsafe questions. It's crucial to apply additional AI safety measures in use cases that require safe and moderated responses.

License

Our OpenChat 3.5 code and models are distributed under the Apache License 2.0.

Dataset Details

OpenChat 3.5 was trained with C-RLFT on a collection of publicly available high-quality instruction data, with a custom processing pipeline. We detail some notable subsets included here:

Citation

@article{wang2023openchat,
  title={OpenChat: Advancing Open-source Language Models with Mixed-Quality Data},
  author={Wang, Guan and Cheng, Sijie and Zhan, Xianyuan and Li, Xiangang and Song, Sen and Liu, Yang},
  journal={arXiv preprint arXiv:2309.11235},
  year={2023}
}

Acknowledgments

We extend our heartfelt gratitude to AutoMeta and caesus from Alignment Lab AI, LDJ and Teknium from Nous Research, alpin and TearGosling from Pygmalion AI for their substantial contributions to data collection and model training.

Special thanks go to Changling Liu from GPT Desk Pte. Ltd., Qiying Yu at Tsinghua University, Baochang Ma, and Hao Wan from 01.AI company for their generous provision of resources. We are also deeply grateful to Jianxiong Li and Peng Li at Tsinghua University for their insightful discussions.

Furthermore, we appreciate the developers behind the following projects for their significant contributions to our research: Mistral, Chain-of-Thought Hub, Llama 2, Self-Instruct, FastChat (Vicuna), Alpaca, and StarCoder. Their work has been instrumental in driving our research forward.

LemTenku
/

chat