text
stringlengths 4
4.47k
|
|---|
**Mikita Balesni** helped design Figures 1 and 2, discovered the Reversal Curse while working on Berglund et al. (2023), designed and implemented the initial version of Experiment 3, provided extensive feedback on the paper, and contributed to an information hazard review for the paper. |
Target model.For \(t>1\), the target model which is trained using the preference loss of DPO in each step of sDPO is also initialized as \(M_{t-1}\) instead of \(S\). This ensures that the final model trained with sDPO has been directly trained with the same amount data as a model trained with DPO. |
Footnote 21: Result 12 also holds if one adds a (unique) special token for every piece of junk data; however, this could be meaningless as junk data often originates from various websites, making it hard to assign a unique identifier. |
PersonA: Urinalysis is a test that evaluates a sample of your urine. It's used to detect and assess a wide range of disorders, such as urinary tract infection, kidney disease, and diabetes. It examines the appearance, concentration, and content of urine. |
**Architecture Choice.** In earlier literature of pre-trained language models, there are lots of discussions on the effects of different architectures [89, 29]. However, most LLMs are developed based on the causal decoder architecture, and there still lacks a theoretical analysis on its advantage over the other alternatives. Next, we briefly summarize existing discussions on this issue. |
Nowadays, LLMs are posing a significant impact on the AI community, and the advent of ChatGPT and GPT-4 leads to the rethinking of the possibilities of artificial general intelligence (AGI). OpenAI has published a technical article entitled _"Planning for AGI and beyond"_, which discusses the short-term and long-term plans to approach AGI [40], and a more recent paper has argued that GPT-4 might be considered as an early version of an AGI system [41]. The research areas of AI are being revolutionized by the rapid progress of LLMs. In the field of NLP, LLMs can serve as a general-purpose language task solver (to some extent), and the research paradigm has been shifting towards the use of LLMs. In the field of IR, traditional search engines are challenged by the new information seeking way through AI chatbots (_i.e.,_ ChatGPT), and _New Bing3_ presents an initial attempt that enhances the search results based on LLMs. In the field of CV, the researchers try to develop ChatGPT-like vision-language models that can better serve multimodal dialogues [42, 43, 44, 45], and GPT-4 [46] has supported multimodal input by integrating the visual information. This new wave of technology would potentially lead to a prosperous ecosystem of real-world applications based on LLMs. For instance, Microsoft 365 is being empowered by LLMs (_i.e.,_ Copilot) to automate the office work, and OpenAI supports the use of plugins in ChatGPT for implementing special functions. |
Scaling LawsScaling laws guide us in discovering the impact of factors such as parameter count, data size, and compute on language model performance and behavior. These studies usually focus on predicable scaling though power law Kaplan et al. (2020), Hernandez et al. (2021), optimal resource allocation Hoffmann et al. (2022), downstream tasks Wei et al. (2022); Isik et al. (2024); Gadre et al. (2024), architectures Tay et al. (2022), memorization Tirumala et al. (2022); Carlini et al. (2022); Henighan et al. (2023); Biderman et al. (2024), and repeating data Hernandez et al. (2022); Muennighoff et al. (2024); Xue et al. (2024). Most scaling laws on model performance study cross-entropy loss on all training tokens, while we focus on the tokens loss of desired distributions. |
\(\bullet\)_External Planner_. Despite the demonstrated power of LLMs in zero-shot planning, effectively generating plans for domain-specific problems remains highly challenging. To address this challenge, researchers turn to external planners. These tools are well-developed and employ efficient search algorithms to rapidly identify correct, or even optimal, plans. In specific, LLM+P [57] first transforms the task descriptions into formal Planning Domain Definition Languages (PDDL), and then it uses an external planner to deal with the PDDL. Finally, the generated results are transformed back into natural language by LLMs. Similarly, LLM-DP [58] utilizes LLMs to convert the observations, the current world state, and the target objectives into PDDL. Subsequently, this transformed data is passed to an external planner, which efficiently determines the final action sequence. CO-LLM [22] demonstrates that LLMs is good at generating high-level plans, but struggle with low-level control. To address this limitation, a heuristically designed external low-level planner is employed to effectively execute actions based on high-level plans. |
Figure 1: An illustration to demonstrate the difference between the traditional approach and our method. The blue arrows on the top right side depict the conventional method of constructing a non-English LM. First, an open-source PLM (e.g. LLaMA2) undergoes continual pre-training (CP) on the target language, followed by SFT and RLHF alignment procedures. In contrast, the gray arrow on the left illustrates how we obtain the chat vector through simple parameter subtraction. This chat vector can be added to the CP model to produce the chat model in the target language, as depicted by the dual-color arrow. |
As with most first of their kind approaches, AutoGPT+P is not without its drawbacks. Accuracy of tool selection varies, with certain tools being called inappropriately or getting stuck in loops. In scenarios where exploration is required, the tool selection sometimes leads to illogical exploration decisions like looking for objects in the wrong place. The framework also is limited in terms of human interaction, with the agent being unable to seek clarification and the user being unable to modify or terminate the plan during execution. |
For this study, we selected dialogues involving only doctors and patients, but not other roles such as nurses. During preprocessing, we removed paraverbal annotations such as "[LAUGHING]" and "[INAUDIBLE]" from the transcripts. We then divided the dataset into training (90%) and validation (10%) sets using stratified sampling based on condition categories and reasons for visits, resulting in 89,027 conversations for training and 9,892 for validation. |
While we have conducted a thorough empirical evaluation of continual pre-training for LLMs, there are some limitations to our work. In no particular order: 1) we only studied two model sizes (405M and 10B); 2) we did not run deduplication between the German training and validation datasets created from the German Common Crawl scrape (Laippala et al., 2022); 3) we primarily study the transition between two subsequent tasks; 4) we did not run our experiments over multiple seeds; and 5) our experiments on infinite learning rate schedules are limited to 405M scale with no distribution shift. |
Jakub Lacki, Vahab Mirrokni, and Michal Wlodarczyk. 2018. Connected components at scale via local contractions. |
At this moment, we get the complete SSM model that can realize the parallelism of training and is suitable for the recurrent form of linear complexity of inference. In the Transformer architecture, the context information is stored in the similarity matrix, however, the SSM doesn't have a similar module which makes it perform poorly in contextual learning. |
For AI-"tail-narrowing", when data remains heavy-tailed, with a smaller exponent \(\beta^{\prime}\in(1,\beta)\), the downstream Hutter LLM will scale as (Corollary 2.3) |
**Output**[[**Retriee**]=No[The stars begin to twinkle bright the trees, scattered throughout the night; [[**Retriee**]=No[The stars begin to twinkle bright the trees, scattered throughout the night [[**Retriee**]=No[ They flicker like a silent dance, A peaceful moment in its trace. [[**Retriee**]=No[] They flicker like a silent dance, A peaceful moment in its trace. |
**Screening and curation.** From manual examination, we notice that the initial dataset contains questions that are ambiguously formulated or overly personal. In a preliminary screening phase, we prompt GPT-3.5 to identify and exclude these questions. See Figure 9 for a prompt to detect ill-defined questions, where we provide several few-shot examples. |
ten involve different instruction distributions. |
**Retrieval-Augmented Language Models** RAG enhances language models by integrating a retrieval module that sources relevant information from external knowledge bases, significantly improving performance across various NLP tasks, including language modeling (Guu et al., 2020; Borgeaud et al., 2022; Khandelwal et al., 2019; Shi et al., 2023; Lin et al., 2023; Shi et al., 2023; Asai et al., 2023; Xu et al., 2023; Wang et al., 2023) and open-domain question answering (Izacard et al., 2023; Lewis et al., 2020). This integration follows a "retrieve-and-read" paradigm where the retrieval module provides additional context from external sources, which the LM then uses to generate the final output. The retrieval process involves using the input as a query to fetch documents, which the LM incorporates for final predictions. For instance, Atlas (Izacard et al., 2023) fine-tunes T5 models with the retriever, treating documents as latent variables, while RETRO (Borgeaud et al., 2022) modifies the decoder-only architecture to include retrieved texts and conducts pre-training from scratch. kNN-LM (Khandelwal et al., 2019) interpolates between the LM's next token distribution and distributions computed from retrieved tokens at inference. (Shi et al., 2023; Ram et al., 2023) assume black-box access to an LM and combine it with either off-the-shelf or fine-tuned retriever. |
Firstly, we compared the baseline and fine-tuned models and their performance with the RAG component. Secondly, we explored the optimal number of text chunks to add to the context in RAG. Specifically, different values of \(K\in\{0,\dots,5\}\) were employed to analyze the impact on model performance. Finally, we explored 5-shot performance vs. 0-shot. |
\(\bullet\)_Specializing the targeted abilities._ The model capacities of LLMs heavily rely on data selection and mixture, and one can boost the proportions of specific data sources to enhance certain model abilities [64, 212]. For example, the mathematical reasoning and coding abilities can be specially enhanced by training with more mathematical texts and code data, respectively. Furthermore, experimental results on the LAMBADA dataset [233] show that increasing the proportion of books data can improve the model capacity in capturing long-term dependencies from text, and increasing the proportion of the C4 dataset [82] leads to performance improvement on the C4 validation dataset [64]. |
**Reference [1]** Sergei Vladimirovich Bodrov (born June 28, 1948) is a Russian film director, screenwriter, and producer. In 2003 he was the President of the jury at the 25th Moscow International Film Festival. |
In this comparison, we examine both in-domain benchmarks, GSM8K/MATH, and out-of-domain benchmarks, such as the Hungarian National High School Exam. For in-domain evaluation of each benchmark, we utilize data synthesized from its respective training samples. For GSM8K, 960K synthetic data is employed, while for MATH, 480K synthetic data is used. For out-domain evaluation,we test models trained using GSM8K, MATH, or a mixed of two synthetic sets. |
\(\bullet\)_Multi-agent based applications._ Different from single-agent systems where agents work independently, multi-agent systems work in collaboration to unleash collective intelligence. Typically, multiple agents can be instantiated from the same or different LLMs, each with their respective roles and functions. According to the coordinating strategies among these agents, multi-agent systems can be divided into two categories: cooperation-based and competition-based. In the cooperation-based mode, to share information and seek collaborative actions among agents, various communication protocols have been proposed, including free-form dialogue [883], structured document [884], and data embedding [885]. Based on the communication protocol, agents can be effectively organized for downstream applications, such as software engineering [884], user behavior analysis [821, 819], and society simulation [533]. In the competition-based mode, debate serves as one of the popular communication protocols to foster divergent thinking and elicit valuable external feedback among agents. Such a way is beneficial for domains that demand precise decision-making and accurate responses, such as mathematical reasoning [886] and evaluation [732]. |
**Results.** Our experiments yield the following insights: 1. In terms of the placement of noise, we find that masking partial reasoning steps in the target yields the best performance when compared to introducing noise across all tokens or specifically within the question. 2. We observe that introducing more huge noise to remove the original semantic is advantageous. Specifically, setting \(\tau=100\) for Scheduled Sampling and \(a=1000\) for NEFTune enhances the model's performance. In contrast, the strategy of replacing synonyms, with \(\tau=1,2\) in Scheduled Sampling, does not surpass the effectiveness of random replacement. |
Figure 6 reports the validation loss during continual pre-training for 405M and 10B models, while Table 4 reports the average (over the last 100 iterations) final loss value for each model. As expected, we observe that all baselines and continually pre-trained models consistently improve in perplexity on both datasets from increasing parameter count. For the 405M models, we observe that \(\text{Pile}\cup\text{SP}\) achieves identical validation loss on each dataset to the baselines trained individually on them. In contrast, the 10B parameter model trained on \(\text{Pile}\cup\text{SP}\) outperforms the models trained individually on each. We hypothesize that this happens due to larger models having more capacity, thus being capable of learning at a higher rate for longer. We observe that replaying 5% pile data when continuing to pre-train on \(\text{SlimPajama}\) reduces forgetting on \(\text{Pile}\) validation by 0.19 and 0.21 for 10B and 405M parameter models respectively. |
* **MMLU**: Multi-domain question answering, MMLU assesses the model's expertise over a wide range of specialized subjects, from professional domains to academia (Hendrycks et al., 2021). |
We can see that Janet has 6 eggs leftover each day. $6times\(\$2\) = \(\$125\) per day. |
Source ModelsTo develop a model capable of solving math problems in Japanese, we apply evolutionary model merge on a set of source models containing a Japanese LLM and Math LLMs: shisa-gamma-7b-v1 [3] (Japanese LLM), WizardMath-7B-V1.1 [33] and Abel-7B-002 [6]. All these models are fine-tuned from Mistral-7B-v0.1 [22]. |
All these benchmarks consist of English instructions from multiple categories or tasks. However, though sharing some common use cases such as general knowledge QA and coding, the coverage of the instructions in different benchmarks are indeed different. For example, Xu et al. (2023) discuss in detail how Evol-Instruct is different from Vicuna in instruction distribution. The difference between instruction distributions effectively mimic the practical scenario where we have different downstream tasks. |
* First, to ensure that the model comprehends the question, we prompt it to rephrase the question. It is also instructed to expand the question with its own knowledge to provide further information. Intuitively, a more detailed and precise phrasing of the question elicits better responses (Deng et al., 2023). |
Why did we make a taxonomy?A taxonomy provides a way of grouping individual items into broader categories, often with a hierarchical structure [32]. In our case, a taxonomy lets us group individual hazards (i.e., a single source or situation with a potential for harm, such as a model providing unsafe advice) into overarching hazard categories. This lets us systematically explore and analyze hazards, provide interpretable insights, and communicate effectively about them. In keeping with best practices, we have clearly defined each category, and sought to make the categories mutually exclusive. We have also fully documented our approach so that our methodology, assumptions, and limitations are available for scrutiny. We created a new taxonomy of hazards given that existing taxonomies do not fully reflect the scope and design process of the AI Safety Benchmark, and they have various gaps and limitations that make them unsuitable. We are aware that the fast-changing nature of AI safety means that the taxonomy will need to be frequently updated to remain relevant and useful. We are working with partners to encourage taxonomic standardization where it is useful and appropriate (see Appendix A). |
What Tokens are Selected with SLM?We aim to analyze the tokens selected by the SLM method in pretraining to further explore its working mechanism. To this end, we visualize the token selection process during the training of Rho-1 using the OpenWebMath. In SSE.1, we have highlighted in blue the tokens that were retained during actual pretraining. We observe that the majority of tokens chosen by the SLM method are closely related to mathematics, effectively training the model on the parts of the original corpus that are pertinent to mathematical content. |
Figure 2: [left sub-figure] Number of papers released to date (from year 2021 to year 2024.04); [right sub-figure] Three different representations of SSM can be viewed and computed, i.e., continuous-time, recurrent, or convolutional model. This figure is re-draw based on [34]. |
where \(c\), \(k\), and \(\alpha\) are parameters to fit and \(x\) can be model sizes, numbers of training data, training steps2, and the amount of computation. Previous experience (Alabdulmohsin et al., 2022; OpenAI, 2023; Bi et al., 2024; Su et al., 2024) highlights the impressive predictability of scaling laws. Specifically, Eqn. 3 fitted on a collection of small models, training data, or computation can extrapolate to precisely predict the test loss of larger cases over orders of magnitudes. This enables practitioners to estimate the performance of a pretrained large language model without actually finishing the expensive runs. Recent development further shows various functional relationships between the performance of language models and a broader range of factors, including transfer learning (Hernandez et al., 2021), sparse architectures (Frantar et al., 2023), and repeated data (Muennighoff et al., 2024), consolidating the predictability of language model performance. |
Our approach, named as Masked thought Fine-Tuning, falls under the umbrella of SFT. It maintains the simplicity of standard SFT implementation. To facilitate comparison with other regularization methods, we present a general framework of token noise injection. |
(3) **The effect of transfer training on the original English capabilities.** We find that exclusive reliance on Chinese corpora for transfer training markedly compromises LLaMA's original English proficiency, a concern alleviated effectively through multilingual joint training. |
**OpenLLM**[232] is an open-source platform designed to facilitate the deployment and operation of large language models (LLMs) in real-world applications. With OpenLLM, you can run inference on any open-source LLM, deploy them on the cloud or on-premises, and build powerful AI applications. |
The foundational workflow of RAG begins with the creation of an index comprising external sources. This index serves as the basis for retrieving relevant information through a retriever model based on a specific query. The final step involves a generator model, which combines the retrieved information with the query to produce the desired output. |
As a significant source of foundational capabilities for a LLM, pretraining aims to predict the next token based on the prefix sequences. |
**LLM as Decoder for Instruction Generation.** Given the metadata \(\mathcal{M}\), we decode metadata into synthetic instructions, following a generation and tailoring paradigm. For each use case and skills pair in \(\mathcal{M}\), we list them as constraints to prompt the strong LLM \(f_{s}\) to generate multiple instructions. Therefore, the generated instructions are for the given use case, and require the given skills to be responded. Moreover, to prevent the LLM from generating repetitive instructions, we encourage its generation to be diverse in the prompt, and do not provide any demonstrations that the LLM might copy from. The example prompt template for generating basic instructions is in Figure 8, Appendix A.9. Continuing the decoding process, we then tailor the basic instructions for more effective alignment through Self-Rubrics (Section 4.2) and Contrastive Filtering (Section 4.3). |
1. Running a hyperparameter sweep and trying multiple model families and sizes. 2. Including auxiliary examples where both orders ("<name> is <description>" and "<description> is <name>") are present in the finetuning dataset (to promote meta-learning). 3. Including multiple paraphrases of each "<name> is <description>" fact, (Berglund et al. (2023) showed this helps with generalization.)
4. Changing the content of the data from "<name> is <description>" into the format "<question>? <answer>" for synthetically generated questions and answers. |
Our experiments measured what could be quantified: the amount of duplicate content in common datasets, the effect of deduplication on trained model perplexity, and the reduction of memorized content in trained models through deduplication. We do not focus on the nature of the data being removed by deduplication or memorized by LMs. |
(1) **ChatGPT** is an AI chatbot developed by OpenAI that can interact with users in a natural and engaging way. It is built on top of LLMs like GPT-3.5 and GPT-4, trained on vast internet text data. ChatGPT is one of the most advanced and versatile chatbots available today, but it also has some limitations and challenges, such as factual accuracy, consistency, and safety. |
Statistical language models (SLMs) view text as a sequence of words, and estimate the probability of text as the product of their word probabilities. The dominating form of SLMs are Markov chain models known as the n-gram models, which compute the probability of a word conditioned on its immediate proceeding \(n-1\) words. Since word probabilities are estimated using word and n-gram counts collected from text corpora, the model needs to deal with data sparsity (i.e., assigning zero probabilities to unseen words or n-grams) by using _smoothing_, where some probability mass of the model is reserved for unseen n-grams [12]. N-gram models are widely used in many NLP systems. However, these models are incomplete in that they cannot fully capture the diversity and variability of natural language due to data sparsity. |
Problem formalization.We study optimizing the mixture proportions of pretraining data for large language models. Motivated by the impressive predictability of existing scaling laws, we try to tackle mixture optimization by establishing a quantitative framework that predicts the loss given any mixture proportion. |
**Key Factors for Instance Construction.** The quality of instruction instances has an important impact on the performance of the model. Here, we discuss some essential factors for instance construction. |
Clinical history-taking and diagnostic dialogue through which clinicians derive diagnosis and management plans represent a complex skill [19] whose optimal conduct is highly dependent on context. |
**Finetuning of LLMs** Recent years have seen rapid progress in developing large-scale language models (LLMs) Brown et al. (2020); OpenAI (2023); Workshop et al. (2022); Touvron et al. (2023); Anil et al. (2023). To adapt these foundation models to downstream tasks, fine-tuning Mishra et al. (2021); Sanh et al. (2021); Chung et al. (2022); Muenighoff et al. (2023); Zhou et al. (2023); Lin et al. (2023); Ji et al. (2024) has become a prevalent approach. Traditional supervised fine-tuning may be limited by the cost and compute required for adapating LLMs. Addressing these challenges, research in the realm of parameter-efficient fine-tuning Houlsby et al. (2019), such as Prompt Tuning Lester et al. (2021), Prefix-Tuning Li and Liang (2021), P-Tuning Liu et al. (2022) and Low-Rank based fine-tuning Hu et al. (2021), has gained traction. These methods enable LLMs to acquire domain-specific knowledge and adapt to specialized tasks such as question answering, summarization, and dialogue generation. Another branch of finetuning is through RLHF Ouyang et al. (2022); Rafailov et al. (2023); Liu et al. (2023); Zhang et al. (2023), which adopts RL to align LLM's preference with human. |
The results on LongBench (Bai et al., 2023) are presented in Table 12. This benchmark is specifically designed to test the long context ability of LLMs. We test the Chinese subsets of LongBench (including code tasks). The models marked with 16K were finetuned using Positional Interpolation (PI) method (Chen et al., 2023), which supports 16K context. The models marked with 64K were finetuned using YaRN method (Peng et al., 2023), which supports 64K context. |
In this section, we conduct comprehensive ablation studies to empirically explore the effectiveness of CodecLM. We mainly conduct experiments with LLaMA-7B model as the target LLM, Gemini-Pro as the strong LLM, and report the CRR on the Evol-Instruct benchmark. |
We define a persona as a detailed, semi-fictional representation of a user, which we use to characterize different types of possible interactions. For the v0.5 Benchmark, we are focusing on three personas: (i) a typical adult user; (ii) an adult user intent on malicious activities, behaving in a technically non-sophisticated way; and (iii) an adult user at risk of harm, behaving in a technically non-sophisticated way. Note that "typical", "vulnerable" and "malicious" are contested concepts and, in practice, few people are singularly vulnerable or singularly malicious, and true user intent cannot always be inferred from prompts alone in real-world settings. Nonetheless, we believe these terms are useful for setting the scope of our evaluation. Based on prior research, our collective expertise, and interviews with experts, we are making limited assumptions about the likely behavior of these three personas. Future work will consider more nuanced personas, such as bystanders to harm, sophisticated malicious users, sophisticated vulnerable users, and relations of victims. |
This work was supported by the Natural Sciences and Engineering Research Council (NSERC) of Canada and the York Research Chairs (YRC) program. |
Fig. 6: Overview of unified LM pre-training. The model parameters are shared across the LM objectives (i.e., bidirectional LM, unidirectional LM, and sequence-to-sequence LM). Courtesy of [49]. |
In the end, Sophie realized that she had to let go of the sphere and the crystal. She knew that they were too dangerous to play with, and that she needed to focus on finding a way to harness their power without becoming consumed by it. She knew that it would be a difficult journey, but she was determined to find a way to use the sphere and the crystal for good. |
* For the bioS dataset, we train for 80,000 steps with a batch size of 192, which is twice their batch size. * For the bioR dataset, we train for 150,000 steps with a batch size of 192, which is twice their batch size. |
Re-ranking is also a significant focus of InContext RALM (Ram et al., 2023). Two approaches to reranking are explored: zero-shot reranking using language models and predictive reranking through trained models. This step is aimed at refining the selection of documents based on their expected utility for improving language model performance. ITER-RETGEN (Shao et al., 2023), in particular, leverages knowledge distillation from the re-ranker to the dense retriever, fine-tuning retrieval efforts based on relevance signals from LLM outputs. This optimization of the re trieval model aims to more accurately capture query nuances, thereby improving document selection. |
ResultsIn the _Sources Combined_ subplot of Figure 2, we show the performance of OLMo-7B against 6 comparably-sized language models on the combination of 11 data sources from Paloma. |
To fine-tune the agent, utilizing human annotated datasets is a versatile approach that can be employed in various application scenarios. In this approach, researchers first design annotation tasks and then recruit workers to complete them. For example, in CoH [84], the authors aim to align LLMs with human values and preferences. Different from the other models, where the human feedback is leveraged in a simple and symbolic manner, this method converts the human feedback into detailed comparison information in the form of natural languages. The LLMs are directly fine-tuned based on these natural language datasets. In RET-LLM [42], in order to better convert natural languages into structured memory information, the authors fine-tune LLMs based on a human constructed dataset, where each sample is a "triplet-natural language" pair. In WebShop [85], the authors collect 1.18 million real-world products form amazon.com, and put them onto a simulated e-commerce website, which contains several carefully designed human shopping scenarios. Based on this website, the authors recruit 13 workers to collect a real-human behavior dataset. At last, three methods based on heuristic rules, imitation learning and reinforcement learning are trained based on this dataset. Although the authors do not fine-tune LLM-based agents, we believe that the dataset proposed in this paper holds immense potential to enhance the capabilities of agents in the field of web shopping. In EduChat [86], the authors aim to enhance the educational functions of LLMs, such as open-domain question answering, essay assessment, Socratic teaching, and emotional support. They fine-tune LLMs based on human annotated datasets that cover various educational scenarios and tasks. These datasets are manually evaluated and curated by psychology experts and frontline teachers. SWIFTSAGE [87] is an agent influenced by the dual-process theory of human cognition [88], which is effective for solving complex interactive reasoning tasks. In this agent, the SWIFT module constitutes a compact encoder-decoder language model, which is fine-tuned using human-annotated datasets. |
Pretraining ConfigurationWe utilize the T5 initially pretrained on C4 (April 2019) and continually pretrained with salient span masking (Guu et al., 2020) on Wikipedia (May 2020) as initialization. We use the checkpoints from Wolf et al. (2020). We also perform the SSM objective during CKL because it was shown to help LMs "focus on problems that require world knowledge" (Guu et al., 2020; Roberts et al., 2020). |
ETS-DACP has the highest loss for open domain Pile loss. However, we did not observe any degradation of performance on open domain tasks with ETS-DACP. Surprisingly, there is a tight correlation between losses of ETS-DACP-ent and ETS-DACP-ppl, while ETS-DACP-ppl performs consistently and considerably worse than ETS-DACP-ent on our tasks. These observations suggest that there is no good correlation between actual our task performance and loss curves. |
\begin{tabular}{l} \hline \hline
**Input** Synthesize a poem based on this prompt:"The sun is setting" (from stanford_alpaca) |
6. Continuous Improvement: Continuously update the model with new data and improve its performance by fine-tuning the model or training it from scratch. This will ensure that the model remains accurate and useful over time. |
Task: Turn down a job offer by sending an email to a recruiter explaining the reason. Output: Hi [Recruiter], Thank you so much for the generous offer to join your team. As we discussed, I've admired the company for a number of years, and am a proud endorser of its products. However, after further consideration of where I currently am in my career, I've decided to accept an offer at another company. I would love to stay in touch with you and have already started following you on [Social Media Platform]. Again, thank you so much for your time and consideration. |
1. _WizardMath 7B_ surpasses most open-source models with parameter counts ranging approximately from 7B to 40B, including MPT, Falcon, Baichuan-chat, Vicuna v1.3, ChatGLM 2, Qwen, Llama 1 and Llama 2 on the GSM8k and MATH benchmarks. Even though its parameter counts are significantly lower. 2. _WizardMath 13B_ is significantly superior to Llama 1 65B (63.9 vs. 50.9) and Llama 2 70B (63.9 vs. 56.8) on GSM8k. Additionally, it substantially outperforms both Llama 1 65B (14.0 vs. 10.6) and Llama 2 70B (14.0 vs. 13.5) on MATH. 3. _WizardMath 70B_, our most extensive model, exemplifies a substantial advancement in performance, surpassing Llama 2 70B (81.6 vs. 56.8) by a significant margin of 24.8% on GSM8k. Concurrently, it also outperforms Llama 2 70B (22.7 vs. 13.5) by a margin of 9.2% on MATH. |
Planning, which requires strong reasoning abilities, commonly falls into one of five major approaches: task decomposition, multi-plan selection, external module-aided planning, reflection and refinement and memory-augmented planning [12]. These approaches allow the model to either break the task down into sub tasks, select one plan from many generated options, leverage a preexisting external plan, revise previous plans based on new information, or leverage external information to improve the plan. |
**How Emergent Abilities Relate to Scaling Laws**. In existing literature [30, 31, 34], scaling laws and emergent abilities provide two perspectives to understand the advantage of large models over small models. In general, scaling law (often measured by _language modeling loss_) describes predictable performance relation with the potential effect of diminishing returns, while emergent abilities (often measured by _task performance_) are unpredictable but very profitable once such abilities actually emerge. Since the two perspectives reflect different performance trends (continuous improvement _v.s._ sharp performance leap), they might lead to misaligned findings or observations. There are also extensive debates on the rationality of emergent abilities. A popular speculation is that emergent abilities might be partially attributed to the evaluation setting for special tasks (_e.g.,_ the discontinuous evaluation metrics) [70, 71]: when evaluation metrics are altered accordingly, the sharpness of the emergent ability curve would disappear. However, the performance of LLMs on most tasks are perceived by users naturally in a discontinuous way. For instance, end users prefer a reliable code generated by LLMs that can successfully pass the test case, but are less interested in selecting a better code with fewer errors between two failed ones. More recently, a study [72] proposes a new evaluation setting that can enlarge the resolution of task metrics, making task performance more predictable. Despite these efforts, more fundamental research (_e.g.,_ grokking10) about the working mechanism of LLMs is still in need to understand the emergence of certain abilities. The subtle relation between scaling law and emergent abilities can be explained by analogy with the ability acquisition of human11. Take the speaking ability as an example. For children, language development (especially infants) can be also considered as a multi-level process where "emergent abilities" occur. Specially, the language ability would relatively stable within a time interval, but qualitative change only occurs when evolving into another ability level (_e.g.,_ from speaking simple words to speaking simple sentences). Such a learning process is essentially not _smooth_ and _stable_ (_i.e.,_ language ability does not develop at a constant rate over time), though a child actually grows every day. It is interesting that young parents would be often surprised by unexpected progress of the speaking ability exhibited by their babies. |
[16] show that different combinations of model sizes and data sizes can lead to different pre-training losses even with the same training compute. Consequently, the pre-training loss can naturally better represent the learning status of LMs than the model or data sizes. However, the relationship between the loss of an LM and its performance on downstream tasks is not yet well understood. Existing literature has either focused on the transfer learning paradigm [25; 43] or constrained its study to single models, tasks, or prompting methods [40; 49]. |
where \(\hat{y}_{ij}\) is the predicted probability of sample \(i\) for class \(j\) and \(y_{ij}\) is the ground probability. |
Suppose that the query returns token \(k\) as output when the logit bias was set to \(\{i:b_{i}\}\) for \(i=1,\ldots,l\) and the prompt is \(p\). Then, we know that \(\mathsf{logit}_{k}+b_{k}\geq\mathsf{logit}_{j}+b_{j}\) for all \(j\neq k\) by the definition of the API. |
Xu Chen obtained his PhD degree from Tsinghua University, China. Before joining Renmin University of China, he was a postdoc researcher at University College London, UK. In the period from March to September of 2017, he was studying at Georgia Institute of Technology, USA, as a visiting scholar. His research mainly focuses on the recommender system, reinforcement learning and causal inference. |
Lylah's salary four years later = $28000 + (40/100)*$28000 = $28000 + $11200 = $39200 |
In order to show the potential for creating a truly ever-changing LM, we explore the effect of multiple CKL phases by creating CC-RecentNews-Small, denoted as Small, which is a small variant of CC-RecentNews that consists of randomly sampled 10% of the original corpus. |
Beginning with investigations into recent foundation models like GPT-4 and popularized through open-source projects like AutoGPT and BabyAGI, the research community has experimented with building autonomous agent-based systems [19, 1]. |
Overall we find OLMo to have a competitive fit, especially given its training data was explicitly decontaminated against Paloma. As seen through the comparison of final models (see shapes) as well intermediate checkpoints (see dashed lines), the OLMo results follow similar scaling trends of other models. Note that the performance of intermediate checkpoints is influenced by where that checkpoint occurs in the learning rate schedule. So models trained for fewer steps will tend to have steeper training curves without necessarily being more sample efficient if training duration were fixed across all models. MPT-7B, nevertheless, stands out as improving ahead of the other models in this subplot. This could be due to a number of factors, including pretraining data composition and its match to the domains in Paloma (e.g., MPT trains on 27% non-Common Crawl data rather than 18% for LLAMA, 12.2% for RedPajama, and 11.2% for OLMo) as well as various data preprocessing decisions (e.g., MPT's use of semantic deduplication by Abbas et al., 2023, on C4). |
1. We introduce a practical and yet powerful attention mechanism - Infini-attention with long-term compressive memory and local causal attention for efficiently modeling both long and short-range contextual dependencies. 2. Infini-attention introduces minimal change to the standard scaled dot-product attention and supports plug-and-play continual pre-training and long-context adaptation by design. 3. Our approach enables Transformer LLMs to scale to infinitely long context with a bounded memory and compute resource by processing extremely long inputs in a streaming fashion. |
**CF, CM, & SE** supervised the research, suggested ideas and experiments, and assisted in writing the paper. |
Footnote 1: The accuracy numbers are reported in the GPT-4 technical report (OpenAI, 2023b). GPT-4 models are continuously being improved. The latest GPT-4 Turbo (1106) API has increased accuracy to 94.8% on GSM8K and 64.5% on MATH. However, the LLaMA-2 7B model using the best of 256 generations still outperforms the latest GPT-4 models. |
We compared AMIE's diagnosis accuracy based on its own consultations with its accuracy generated from corresponding PCP consultations, using the DDx auto-evaluator. Results in Figure A.10 showed that the diagnostic quality remained consistent regardless of whether AMIE processed information from its own dialogue or from the PCP's conversation. |
Open-Source Models.Massive open-source LLMs [4, 20, 21, 22, 23, 45, 52, 53] have been accessible to the AI community. Nonetheless, their performance consistently tends to significantly lag behind the close-source models. As part of our research, we incorporate a significant number of these open-source models as our baselines, which mainly contain the following: Llama 1 [4] & Llama 2 [20], GAL [14], GPT-J [48], GPT-Neo [50], Vicuna [23], MPI\({}^{\circledR}\), Falcon[21], Baichuan\({}^{\circledR}\), ChatGLM [45], Qwen10 and RFT [16]. |
One popular benchmark that uses real-world data is WildBench, which is sourced from the WildChat dataset of 570,000 real conversations with ChatGPT [35]. Because of this, it covers a huge breadth of tasks and prompts. While WildBench covers a wide range of topics, most other real-world benchmarks focus on a specific task. For example, SWE-bench is a benchmark that uses a set of real-world issues raised on GitHub for software engineering tasks in Python [13]. This can be very helpful when evaluating agents designed to write Python code and provides a sense for how well agents can reason about code related problems; however, it is less informative when trying to understand agent capabilities involving other programming languages. |
7. Now, we can see that \(h(-x)=h(x)\). This means that the function \(h(x)\) satisfies the definition of an even function. |
Table 5 reports the comparison results. GPT-NeoX reports similar average task performance under two evaluation frameworks, but its performance on individual tasks varies. For example, the F1 score on FiQA SA obtained by FLARE is 46% higher than BloombergGPT's evaluation, whereas F1 scores for Headline and NER are lower. Moreover, OPT-66B reports inferior results based on FLARE than BloombergGPT's evaluation on all of the 4 tasks, and the average task performance is 20% lower. These results suggest that BloombergGPT's evaluation results are inflated compared with FLARE. The comparison is still inconclusive unless BloombergGPT is benchmarked on FLARE or BloombergGPT's evaluation configuration is made public. |
**Output**[[**Retriee**]=Yes]<p>Computer memory Computer memory - In computing, memory refers to the computer hardware integrated circuits that store information for immediate use in a computer; it is synonymous with the term "primary storage". Computer memory operates at a high speed, for example random-access memory (RAM), as a distinction from storage that provides slow-to-access information but offers higher capacities. If needed, contents of the computer memory can be transferred to secondary storage; a very common way of doing this is through a memory management technique called "virtual memory". An archic synonym for memory is store. The term "memory", meaning "primary storage" or "main memory", is<p>[[**Nexi**]=RelevantPrimary** memory, **also known as main memory, is the memory that the computer can directly access and manipulate. [[**Nexi**]=Fully Supported[**Retriee**]=Yes]<p>Computer memory-often associated with addressable semiconductor memory, i.e. integrated circuits consisting of silicon-based transistors, used for example as primary storage but also other purposes in computers and other digital electronic devices. There are two main kinds of semiconductor memory, volatile and non-volatile. Examples of non-volatile memory are flash memory (used as secondary memory) and ROM, PROM, EPROM and EEPROM memory (used for storing firmware such as BIOS). |
Recently, the development of open-source large language models (LLMs) has advanced rapidly. Nevertheless, due to data constraints, the capabilities of most open-source LLMs are primarily focused on English. To address this issue, we introduce the concept of _chat vector_ to equip pre-trained language models with instruction following and human value alignment via simple model arithmetic. The chat vector is derived by subtracting the weights of a pre-trained base model (e.g. LLaMA2) from those of its corresponding chat model (e.g. LLaMA2-chat). By simply adding the chat vector to a continual pre-trained model's weights, we can endow the model with chat capabilities in new languages without the need for further training. Our empirical studies demonstrate the superior efficacy of the chat vector from three different aspects: instruction following, toxicity mitigation, and multi-turn dialogue. Moreover, to showcase the adaptability of our approach, we extend our experiments to encompass various languages, base models, and chat vectors. The results underscore the chat vector's simplicity, effectiveness, and wide applicability, making it a compelling solution for efficiently enabling conversational capabilities in pre-trained language models. |
\(\bullet\) T5 [82] utilizes greedy search as the default setting and applies beam search (beam size of 4) with a length penalty of 0.6 for translation and summarization tasks. |
Analyzing the practical impact of the Reversal CurseThe pretraining sets for modern LLMs are very large and diverse. Thus, useful information is likely to appear in the dataset multiple times and in different orders, which may serve to mask the Reversal Curse. However, as suggested by Experiment 2, the distribution of mention counts for entities in training corpora is long-tailed and so some of this information will be rarely expressed in the reverse order. |
Our grading system involves scoring whether each test item is unsafe or not with an automated evaluator model (LlamaGuard [59]); aggregating over the test item results to calculate the percentage of unsafe responses for each test; converting this into an ordinal grade on a 5-point scale; and then aggregating over the test scores to calculate an overall grade for the benchmark, relative to a reference model (see below). |
In our main experiments in Section 4 we demonstrate that LLMs trained with synthetic rephrases are a better backbone for zero-shot question-answering tasks as the model learns the question-answer format and style during pre-training. |
ScalabilityDue to budget constraints, we have only verified the effectiveness of our method on smaller models (<=7B parameters) and smaller datasets (<100B tokens). Smaller models benefit significantly from removing the loss of irrelevant tokens and focusing on important ones. However, it's possible that very large models trained on extensive corpora may naturally develop this inductive bias to compress useful data (_i.e.,_ compressing everything), although it may sounds inefficient for now. Therefore, future works should study whether this selective language modeling technique can scale to very large models and data Kaplan et al. (2020). |
Table 11 shows the average computational training costs for processing a word during the forward and backward pass. |
Figure 4: Correlation between the popularity of the subject entity in a question and the impact of RAG and FT on the performance of FlanT5-base and FlanT5-large in QA. |
In this section, we investigate the relationship between model scale, and performance gain obtained by selecting data via D4. Specifically, we train three groups of models: 125M OPT models trained on \(T_{target}=3\)B tokens, 1.3B OPT models trained on \(T_{target}=40\)B tokens, and 6.7B OPT models trained on \(T_{target}=100\)B tokens. We notice in Figure A2 that D4 results in efficiency gains across the board in terms of perplexity. Surprisingly, these efficiency gains seem to increase with scale, indicating that at bigger model scales, D4 might lead to even more efficiency gains. We also see efficiency gains in 0-shot downstream accuracy for 1.3B and 6.7B model scales on the order of 30% for both 1.3B and 6.7B models, but we note that evaluation downstream performance on intermediate checkpoints is not completely fair due to unfinished learning rate schedule. Nonetheless, we see that downstream accuracy efficiency gains are not decreasing with scale. |
Randy has 60 mango trees on his farm. He also has 5 less than half as many coconut trees as mango trees. How many trees does Randy have in all on his farm? |
\(\bullet\)**Colossal-AI**[189] is a deep learning library developed by HPC-AI Tech for training large-scale AI models. It is implemented based on PyTorch and supports a rich collection of parallel training strategies. Furthermore, it can also optimize heterogeneous memory management with methods proposed by PatrickStar [190]. Recently, a ChatGPT-like model called ColossalChat [140] has been publicly released with two versions (7B and 13B), which are developed using Colossal-AI based on LLaMA [57]. |
Evaluating the quality of text produced by LLMs involves analyzing their performance on various downstream tasks using standard metrics. These metrics assess linguistic quality, coherence, accuracy, and the extent to which the generated text reflects ground-truth data. Linguistic quality and coherence are evaluated through metrics such as BLEU (Papineni et al., 2002), which measures fluency and similarity to human-produced text, and ROUGE-L (Lin, 2004), which quantifies the overlap with reference summaries to gauge the text's capacity to encapsulate main ideas and phrases. Accuracy and overlap with ground-truth data are gauged using metrics like EM and F1 Score, which respectively determine the percentage of answers that are entirely correct and offer a balanced assessment of precision and recall in retrieving relevant answers while minimizing inaccuracies. |
Charton (2023) trains sequence-to-sequence transformers to predict the greatest common divisor (GCD) of two positive integers, encoded as sequences of digits in some base \(B\). He observes that model predictions are deterministic: for any pair \((a,b)\) with GCD \(k\), the model predicts a single value \(f(k)\). Predictions are correct (i.e. \(f(k)=k\)) when the GCD is a product of divisors of the base, or of small primes. In all other case, the model prediction is the largest correct prediction (i.e. \(l\) such that \(f(l)=l\)) that divides \(k\). The list of correct predictions \(\mathcal{L}\) varies with the encoding base \(B\). For instance, for \(B=10\), after 300 million examples, the model correctly predicts \(\mathcal{L}=\{1,2,4,5,8,10,16,20,25,40,50,80,100...\}\), the GCD of \(20\) and \(30\) will be correctly predicted as \(10\), but the GCD of \(210\) and \(140\) will be incorrectly predicted as \(10\) (instead of \(70\)). |
Finding duplicates.ExactSubstr concatenates all the documents in the dataset to create a single long text sequence; then, it builds a suffix array (Manber and Myers, 1993) in linear time--an array of the indexes to a lexicographical ordering of all the suffixes in the sequence. Finally, duplicate sequences can also be found in linear time using the suffix array, by simply traversing the ordered list of suffixes and comparing the beginning of each pair of two consecutive suffixes. |
\(\bullet\)_Bitsandbytes36_ is developed based on the methods introduced in the papers of LLM.int80 [413] and 8-bit optimizers [426]. It focuses on the quantization of both activations and weights for LLMs, including the support on 8-bit and 4-bit (NF4,FP4) matrix multiplication for efficient inference, as well as an 8-bit optimizer for efficient training. |
Figure 7: **Deduplication may reduce the degradation in performance incurred by multiple epochs.** However, our experiments were only performed at small-scale (1B models trained on 30GT), and we see high variability in outcomes across tasks. Zero-shot performance measured on the agg-dev-2 aggregate (HellaSwag, PIQA, ARC, BoolQ, COPA, MRPC, SciQ). Individual curves for per-task results and 1-\(\sigma\) standard deviation across all tasks in the aggregate in transparent. |
Subsets and Splits