text
stringlengths 4
4.47k
|
|---|
* Natural Questions (NQ) (Kwiatkowski et al., 2019), Trivia QA (Joshi et al., 2017) and HotpotQA (Yang et al., 2018) are the open-domain question-answers based on Wikipedia, mainly focused on common knowledge (e.g., movies, sports, etc). * HuggingFace, Torch Hub, and TensorFlow Hub are from the APIBench (Patil et al., 2023) proposed in the Gorilla paper. These benchmarks measure how to generate the correct, functional, and executable API calls based on the documentation. * PubMed QA (Jin et al., 2019) is a question-answering dataset tailored only for biomedical-research question-answering. It mainly focuses on answering medical and biology questions based on a given set of documents. |
Analysis of standard benchmarksWe define that standard benchmarks include PIQA, SIQA, HellaSwag, WinoGrande, ARC Easy and Challenge, OpenBookQA, CommonsenseQA, BoolQ, and MMLU. This categorization is predicated on the observation that these benchmarks, to varying degrees, evaluate the models' understanding of real-world knowledge, reading comprehension, and commonsense reasoning. The descriptions and observations on these datasets are as follows. |
**Lemma G.1**.: _There exists a positive semidefinite \(\mathbf{A}\in\mathbb{R}^{h\times h}\) such that \(x_{i}^{\top}\mathbf{A}x_{i}=1\) for all \(i\)._Because \(\mathbf{A}\) is positive semidefinite, we can write \(\mathbf{A}=\mathbf{M}^{\top}\cdot\mathbf{M}\) for some \(\mathbf{M}\). |
* For \(N=10K\), we use \(wd=0.02\), \(lr=0.001/0.002\), and batch size 24 with fp16;
* For \(N=20K\), we use \(wd=0.02\), \(lr=0.001/0.002\), and batch size 48 with fp16;
* For \(N=50K\), we use \(wd=0.02\), \(lr=0.001/0.002\), and batch size 96 with fp16;
* For \(N=100K,200K\), we use \(wd=0.02\), \(lr=0.001/0.002\), batch size 192 with fp16;
* For \(N=500K,1M\), we use \(wd=0.01\), \(lr=0.0005/0.001\), batch size 192 with fp16;
* For \(N=2M\), we use \(wd=0.005\), \(lr=0.002\), and batch size 1536 with fp16;
* For \(N=5M\), we use \(wd=0.002\), \(lr=0.0005\), and batch size 1536 with fp16;
* For \(N=10M\), we use \(wd=0.001\), \(lr=0.0005\), and batch size 1536 with fp16. |
\(N>2\) **settings** - We also consider the following settings with more dataset transitions to investigate how well the methods considered scale with more datasets: |
Autonomous agents have long been a research focus in academic and industry communities. Previous research often focuses on training agents with limited knowledge within isolated environments, which diverges significantly from human learning processes, and makes the agents hard to achieve human-like decisions. Recently, through the acquisition of vast amounts of web knowledge, large language models (LLMs) have shown potential in human-level intelligence, leading to a surge in research on LLM-based autonomous agents. In this paper, we present a comprehensive survey of these studies, delivering a systematic review of LLM-based autonomous agents from a holistic perspective. We first discuss the construction of LLM-based autonomous agents, proposing a unified framework that encompasses much of previous work. Then, we present a overview of the diverse applications of LLM-based autonomous agents in social science, natural science, and engineering. Finally, we delve into the evaluation strategies commonly used for LLM-based autonomous agents. Based on the previous studies, we also present several challenges and future directions in this field. |
where \(f(L)\) is a monotonically decreasing function of \(L\), \(\eta\) is the threshold, and the normalized performance of random guess is 0. |
**No fine-tuning and no retrieval.** We evaluate our optimal system without any news retrieval and using the base GPT-4-1106-Preview model. The ablated system attains a Brier score of.206. |
Essentially, scaling laws [14] reveal how to predict the _validation/test_ loss3 of a given model, which can be the final loss when the model is trained to converge, or the loss at a certain step amid training. Scaling laws have been proven to be widely valid across a variety of likelihood-based tasks [15]. By adhering to scaling laws, researchers can uncover patterns in how changes in model parameters and training data impact the overall effectiveness of large language models before actually training them. |
Math.Recent advancements in mathematical reasoning for language models (LMs) have led to the development of various approaches to improve performance on math-related tasks. One approach is to train on math-targeted pre-training data, such as Minerva (Lewkowycz et al., 2022), Llemina (Azerbayev et al., 2023), and DeepSeeKMath (Shao et al., 2024). Another mainstream method is to generate synthetic questions and answers to imitate the training or validation set of target benchmarks. For instance, WizardMath (Luo et al., 2023) leverages a series of operations to increase the complexity of questions and answers using GPT-3.5, while MetaMath (Yu et al., 2023) bootstraps the questions in MATH and GSM8K by rewriting them in different ways, such as semantic rephrasing, self-verification, and backward reasoning. GAIR-Abel (Chern et al., 2023) found that the format of the augmented answers is crucial to final performance, with answers that begin with a paraphrasing of the question followed by a step-by-step solution showing better performance than those in vanilla format. Xwin-Math (Li et al., 2024) further scaled up synthetic SFT data to one million examples and found that the LLaMA-2 7B model (Touvron et al., 2023) can still benefit from data scaling. MMIQC (Liu and Yao, 2024) composed a bundle of datasets that infuse SFT style data (via question-answer rephrasing or directly taken from MetaMath) with a subset of high-quality mathematical pre-training data, such as OpenWebMath (Paster et al., 2023). |
The concurrent work by Allen-Zhu & Li (2023) investigates a related set of failures and potential solutions. Exploring the capability to answer questions based on synthetic biographies, they examine several data augmentation strategies, including incorporating instruction tuning data into pre-training, generating multiple unique biography entries, permuting biography sentences, and substituting pronouns or partial names with full names. They discover that augmentation during the pre-training phase is essential for enhancing downstream question answering performance across various tasks. However, in real pre-training data, some augmentations may not be feasible -- for instance, permuting sentences could degrade language model quality, and it remains uncertain how to best rewrite data during augmentation. Reverse training addresses this issue by presenting a distinct language task (the reversed language) to the language model, thereby avoiding interference with the primary task of left-to-right natural language modeling. |
Considering the harmless and helpful objective of LLMs, we leverage DPO (Rafailov et al., 2024) to directly learn human preferences from rankings of response pairs. |
The parameterized knowledge in the model is fixed at the end of the pre-training, which leads to its shortcomings in terms of knowledge capacity and knowledge updating. While introducing external knowledge can alleviate this to some extent, it remains an imperfect solution. To fundamentally tackle this issue, continual learning Lange et al. (2022); Wang et al. (2023) stands as a promising avenue for future research endeavors. |
**Language Modeling.** As the most fundamental ability of LLMs, _language modeling_ aims to predict the next token based on the previous tokens [1], which mainly focuses on the capacity of basic language understanding and generation. For evaluating such an ability, typical language modeling datasets that existing work uses include Penn Treebank [538], WikiText-103 [539], and the Pile [161], where the metric of _perplexity_ is commonly used for evaluating the model performance under the zero-shot setting. Empirical studies [93, 55] show that LLMs bring substantial performance gains over the previous state-of-the-art methods on these evaluation datasets. To better test the modeling capacity of long-range dependencies in text, the LAMBADA dataset [233] has been introduced, where LLMs are required to predict the last word of sentences based on a paragraph of context. Then, the accuracy and perplexity of the predicted last words are employed to evaluate LLMs. As shown in existing work, the performance on the language modeling tasks typically follows the scaling law [30], which means that scaling language models would improve the accuracy and reduce the perplexity. |
Figure 6: **Is re-phrasing same as any augmentation?** We compare perplexity on the Pile for different augmemntation strategies. 350M parameter models are trained for a total of 15B tokens. **WRAP** (Medium + C4) performs significantly better than traditional augmentations. |
After introducing existing benchmarks, in this part, we will review existing evaluation approaches for assessing the performance of LLMs. To organize our discussion, we categorize LLMs into three different types: _base LLMs_ (pre-trained model checkpoints), _fine-tuned LLMs_ (instruction or alignment fine-tuned model checkpoints), and _specialized LLMs_ (adapted model checkpoints for some specific task or domain). Here, we keep both fine-tuned LLMs and specialized LLMs, to distinguish the different purposes of LLMs: general or specific task solvers. To evaluate the three types of LLMs, we can test the LLM's performance related to different abilities (_e.g._, basic or advanced abilities as discussed in Section 7.1 and 7.2). In general, there are three main approaches to evaluating LLMs, namely benchmark-based approach [364], human-based approach [727], and model-based approach [729]. Table 15 shows an illustration of the relationship among LLM type, evaluation approach, and tested abilities. Next, we will discuss the evaluation approaches for different types of LLMs. |
In information retrieval, the quality of search results is typically evaluated using standard metrics such as Mean Average Precision (MAP), Precision, Reciprocal Rank, and Normalized Discounted Cumulative Gain (NDCG) (Radlinski and Craswell, 2010; Reimers and Gurevych, 2019; Nogueira et al., 2019). These metrics primarily assess the relevance of retrieved documents to a given query. |
Given a set of factual questions, there exists some text corpus containing information that is relevant to these questions. The central assumption of knowledge injection is that given full access to this corpus, it could serve as an auxiliary knowledge base and improve the model's performance on this set of questions. |
Reasoning Framework for Improving Mathematical CapabilityChain-of-thoughts Wei et al. (2022) encourages the LLMs perform multi-step reasoning by specific designed prompts and can improve reasoning performance. Based on this work, many subsequent works suggesting further improvements Fu et al. (2022); Zhang et al. (2022); Kojima et al. (2022). The above works focus primarily on how to improve performance through better prompt design or inference strategies without fine-tuning the model, whereas our work focuses on how to improve the model itself, and thus these approaches are complementary to ours. |
- The base rate for such high-profile resignations in the face of uncertainty is not well-established, which adds to the difficulty in predicting this event. |
Because of these fundamental differences, instead of contemporary CL methods mentioned above, we explore methodologies from the literature that are suitable for CKL (Chen et al., 2020; He et al., 2021; Hu et al., 2021; Wang et al., 2021), modifying and adapting each method according to our needs as CKL methods. Lastly, while it has been pointed out that some of the traditional CL formulations may have little practical importance in real-world scenarios by Prabhu et al. (2020), CKL is much closer to the initial motivation behind CL, which is that the "fundamental characteristic of natural intelligence is its ability to continually learn new knowledge while updating information about the old ones" (Prabhu et al., 2020). Details of related works regarding the traditional CL methods and how CKL methods address the fundamental differences are provided in Appendix A. |
Data.We train Nemotron-4 15B on a pre-training dataset consisting of 8 trillion tokens. At a high-level, the data blend is split into three different types of data: English natural language data (70%), multilingual natural language data (15%), and source-code data (15%). |
Mitigating hallucination.Many widely-used language models utilize supervised finetuning (SFT) to learn to align their interactions with users (Wang et al., 2022b; Zhang et al., 2023b). In particular, there exist many methods of generating synthetic SFT data that can improve capabilities such as reasoning and alignment (Wei et al., 2023a,b). It has been shown, however, that these synthetic data can induce hallucinations into language models by containing nontrivial amounts of hallucinated answers or by forcing models to learn to answer questions that they do not know the answer to (Zhang et al., 2023c). These cases demonstrate that synthetic data, if not applied correctly, can actually increase hallucinations in language models. |
Anisotropy is a common form of representation degeneration that has been observed among various small language models. It consists in reduced angular variability of the representation distribution at a given layer. Previous works (Ethayarajh, 2019; Godey et al., 2024) notice that almost all layers of small Transformers language models are anisotropic. |
- The pressure from political opponents and the recent loss of the Speaker position may contribute to a decision to resign, especially if he feels his influence is waning. |
In addition to the strong results, we have also gleaned insights into the effectiveness of our approach: 1) As the scale of SFT data increases, the model's accuracy tends to plateau when utilizing 256 attempts; however, there is a marked increase using 1 response. This indicates that while the model's upper capability limit remains fairly constant, the performance gains are primarily due to enhanced stability in generating correct answers. 2) The accuracy of solving math problems follows a power law with respect to the number of chain-of-thought (CoT) steps with different SFT data quantities. An expanded SFT dataset improves the reliability of each reasoning step. Further increasing the proportion of training samples with longer CoT steps through resampling can significantly improve the accuracy of the model for difficult questions. 3) An analysis of error types during the scaling process reveals that calculation errors are more readily mitigated compared to reasoning errors. |
Footnote 24: Model parallelism is a more broader term that includes tensor parallelism and pipeline parallelism in some work [75]. |
Footnote 1: All resources are publicly available at [https://github.com/tLabruna/Adapt-LLM](https://github.com/tLabruna/Adapt-LLM). |
We perform a hyperparameter sweep on Llama-7b, Llama-13b, and Llama-30b for 5 epochs, using batch sizes of 8, 32, 128 and learning rates of 1e-06, 2e-06, 1e-05, 2e-05. We use Adam as our optimizer and DeepSpeed level 3 for memory efficiency. We perform full finetuning and do not use any parameter efficient finetuning techniques. We chose these batch sizes to be relatively low. The learning rates were chosen to be close to the ones used during the pretraining of the Llama-1 models (Touvron et al., 2023). The results for Llama-7b are shown in Figure 10. |
- Concerns about format or moderators: Moderate to High. Trump has previously expressed dissatisfaction with debate arrangements, which could influence his decision. |
Proof.: Consider the function \(h(z):=ze^{-Tz}\) for \(z\in(0,1)\). Its derivative is \(h^{\prime}(z)=e^{-Tz}(1-Tz)\). Thus, \(h\) is increasing on \((0,1/T)\) and decreasing on \((1/T,\infty)\). Furthermore, note that \(p_{i}\leq 1/T\) iff \(i\geq T^{1/\beta}\). |
Figure 1: The percentage of pass rates on the HumanEval (164 problems) with a single attempt. All baseline scores are retrieved from the LLM-Humaneval-Benchmarks [39]. Our _WizardCoder_ generates an answer with greedy decoding. |
Fig. 27: Subsequent stages of Macrodata Refinement remove nearly 90% of the documents originally in CommonCrawl. Courtesy of [124]. |
Filtering and Postprocessing.To encourage diversity, a new instruction is added to the task pool only when its ROUGE-L similarity with any exist ing instruction is less than 0.7. We also exclude instructions that contain some specific keywords (e.g., image, picture, graph) that usually can not be processed by LMs. When generating new instances for each instruction, we filter out instances that are exactly the same or those with the same input but different outputs. Invalid generations are identified and filtered out based on heuristics (e.g., instruction is too long or too short, instance output is a repetition of the input). |
Forward-looking Active Retrieval Augmented Generation (FLARE) [168] enhances the capabilities of Large Language Models (LLMs) by iteratively combining prediction and information retrieval. FLARE represents an evolution in the use of retrieval-augmented generation, aimed at improving the accuracy and relevance of LLM responses. |
Figure 3 presents the results with different numbers of relevant rules included in ground truth proofs, where the problem does not contain distracting rules, and the shuffled accuracy is the aggregation of |
For downstream accuracy, we chose to match the exact downstream evaluation done in Zhang et al. [59] since we use OPT architecture and hyperparameters. Similar to Zhang et al. [59], we notice considerable variability across the 16 NLP tasks in Figure A3, motivating us to look at the mean downstream accuracy across tasks. |
In this paper, we introduce SaulLM-7B, a large language model (LLM) tailored for the legal domain. With 7 billion parameters, SaulLM-7B is the first LLM designed explicitly for legal text comprehension and generation. Leveraging the Mistrat 7B architecture as its foundation, SaulLM-7B is trained on an English legal corpus of over 30 billion tokens. SaulLM-7B exhibits state-of-the-art proficiency in understanding and processing legal documents. Additionally, we present a novel instructional fine-tuning method that leverages legal datasets to further enhance SaulLM-7B's performance in legal tasks. SaulLM-7B is released under the MIT License. |
\(\bullet\)_Training on code data._ A major limitation of the original GPT-3 model (pre-trained on plain text) lies in the lack of the reasoning ability on complex tasks, _e.g._, completing the code and solving math problems. To enhance this ability, Codex [105] was introduced by OpenAI in July 2021, which was a GPT model fine-tuned on a large corpus of GitHub code. It demonstrated that Codex can solve very difficult programming problems, and also lead to a significant performance improvement in solving math problems [126]. Further, a contrastive approach [127] to training text and code embedding was reported in January 2022, which was shown to improve a series of related tasks (_i.e._, linear-probe classification, text search and code search). Actually, the GPT-3.5 models are developed based on a code-based GPT model (_i.e._, code-davinci-002), which indicates that training on code data is a very useful practice to improve the model capacity of GPT models, especially the reasoning ability. Furthermore, there is also a speculation that training on code data can greatly increase the chain-of-thought prompting abilities of LLMs [47], while it is still worth further investigation with more thorough verification. |
semantically equivalent references with diverse expressions. Further, LLMs are widely employed as the evaluators of text generation in a reference-free manner, including evaluating a single prediction [631, 642, 632] or comparing several candidates [643, 644, 138, 645]. Nevertheless, LLMs may expose bias (_e.g.,_ order bias or preference for LLM-generated texts over human-written texts) as language generation evaluators, demonstrating disparities when compared to human evaluation [646, 632, 647]. |
In this part, we review some of the popular approaches used for more cost-friendly (and compute-friendly) training and usage of LLMs. |
Brown et al. (2020) finds that large-scale language models have excellent in-context learning (ICL) ability. ICL incorporates input-output demonstrations into the prompt text. With ICL, off-the-shelf LLMs can be employed without additional fine-tuning while achieving comparable performance. Nevertheless, this end-to-end approach tends to underperform when faced with complex reasoning tasks. |
Subsequently, a rigorous cleaning process was utilized to transform the data from raw subsections to clean chunks. This step was done with the "wikiextractor" tool (Attardi, 2015). The division into small, clean (e.g., remove HTML, URLs, etc.) chunks was aimed at enhancing the evaluation of the LLMs' understanding across various knowledge domains and aiding the LLMs in the fine-tuning process. |
Koala [64] is yet another instruction-following language model built on LLaMA, but with a specific focus on interaction data that include user inputs and responses generated by highly capable closed-source chat models such as ChatGPT. The Koala-13B model performs competitively with state-of-the-art chat models according to human evaluation based on real-world user prompts. |
hence the shortest path is an upper bound on \(\mathsf{logit}_{i}\). To prove feasibility, we claim that setting \(\mathsf{logit}_{i}\) to be the distance from \(0\) to \(i\) satisfies all the inequalities. Assume some inequality \(\mathsf{logit}_{i}-\mathsf{logit}_{j}\leq c_{ji}\) is violated. Then we can go from \(0\to j\to i\) in \(G\) with a total weight of \(\mathsf{logit}_{j}+c_{ji}<\mathsf{logit}_{i}\), which contradicts the assumption that \(\mathsf{logit}_{i}\) is the distance from \(0\) to \(i\). |
**LLM-Eval**[13]: LLM-Eval is a manually constructed benchmark for instruction-following evaluation. It has 453 instruction tasks from 17 major categories, including factual question answering, reading comprehension, frame generation, paragraph rewriting, summarizing, math problem solving, reasoning, poetry generation, programming, and more. |
The evolution of operating systems (OS) has unfolded in a progressive way, evolving from rudimentary systems to the complex and interactive OS of today. Initially, operating systems served to bridge the gap between the user-level tasks and the binary functionality of computer hardware, such as electron and gate manipulation. Their evolution saw a transition from simple batch job processing [23] to more advanced process management techniques like time-sharing [24] and multi-task processing [25, 26], which facilitated the handling of increasingly complex tasks. The progress moved toward modularization within the OS, delineating specific responsibilities such as process scheduling [27, 28], memory management [29, 30], and filesystem management [31, 32], enhancing efficiency and manageability. The further advent of graphical user interfaces (GUIs), e.g., Macintosh1, Windows2 and GNOME3, makes operating systems more interactive and user-centric. Meanwhile, the operating system ecosystem has also expanded, offering a comprehensive suite of developer tools (OS SDKs) and runtime libraries. These tools enable application developers to design, implement, and run their applications efficiently within the OS environment [33]. Notable examples of OS ecosystems include Android Studio4, XCode5 and Cloud SDK6. In these ecosystems, the OS provides numerous resources to facilitate software development and serves as a platform for deploying and hosting software applications, leading to a thriving OS-application ecosystem. Nowadays, we stand at the transformative phase to see the potential of intelligent operating systems. With the incorporation of large language models (LLMs), These advanced systems promise to further narrow the communication gap between humans and machines, forwarding a new era of user-computer interaction. |
Figure 2: **Left.** The frontier of expected reward vs KL to the reference policy. DPO provides the highest expected reward for all KL values, demonstrating the quality of the optimization. **Right.** TL;DR summarization win rates vs. human-written summaries, using GPT-4 as evaluator. DPO exceeds PPO’s best-case performance on summarization, while being more robust to changes in the sampling temperature. |
3. The story's plot focuses on Iris Winnow, an employee at the Oath Gazette newspaper in the city of Oath. |
**C-Eval**[13]: C-Eval is a Chinese evaluation suite with 13948 exam questions across 52 subjects and 4 difficulty levels from middle school to professional exams. It includes STEM, humanities, social science and other topics. C-Eval HARD is a subset of 8 challenging math and science subjects requiring advanced reasoning. |
Our approach to pretraining is to train a standard dense transformer architecture on a heavily engineered large pretraining corpora, where our underlying assumption is that when trained on extensive data of high-enough quality, a standard architecture can exhibit advanced capability. This is to say, we may not need much architectural modification, although we have indeed conducted extensive preliminary architectural experiments. In the following subsections, we first detail our data engineering pipeline, then briefly discuss the model architecture. |
\end{table}
Table 7: **German LM Eval Performance of Models Merged with TIES v.s. Continual Pre-training. Normalized accuracy is reported for HellaSwag and exact match (EM) is reported for NaturalQuestions and TriviaQA. All other tasks report unnormalized accuracy. |
Baselines.As we are the first work to consider inverting text directly from language model probabilities, there are no prior baselines to compare to. |
In heterogeneous combination experiment, GPT-3.5-Turbo is used for generating the intermediate answer with \(K=8\), and GPT-4 is then employed to solve for the final answer with \(K=32\). In Figure 7 (right), compared with the result of GPT-4 with \(K=32\), the hierarchical method improves performance from \(35\%\) to \(47\%\), suggesting the deployment of different LLMs at the corresponding level of problem-solving can improve the performance in a cost-effective manner. |
Chain-of-Thought (CoT) prompting [502, 33] is an improved prompting strategy to boost the performance of LLMs on complex reasoning tasks, such as arithmetic reasoning [503], commonsense reasoning [504], and symbolic reasoning [33]. Instead of simply constructing the prompts with input-output pairs like ICL, CoT prompting further incorporates intermediate reasoning steps, which serve as the bridge between inputs and outputs. Figure 14 presents an illustration of CoT. In the following part, we will first elaborate on the basic CoT prompting approach and its improved strategies, then discuss when and why CoT prompting works. |
**Query cost.** This attack reveals the value of 4 logits with each query to the model (the fifth being used as a reference point), for a cost of \(1/4\) queries per logit. |
Figure 7: **Impact of style of synthetic rephrases:** Perplexity across all domains of the Pile comparing different styles of synthetic data. We train 128M parameter models for 3B tokens. |
Overall, the LLaMA-7B model, both with and without the RAG, performs poorly due to its answering style not aligning with the ground truth. By applying domain specific tuning, we significantly enhance its performance. This process enables the model to learn and adopt the appropriate style of answering. However, introducing RAG to a domain-specifically fine-tuned (DSF) model doesn't invariably lead to better outcomes. This might indicate that the model lacks training in context processing and extracting useful information from it. By incorporating our method, RAFT, we train the model not only to match its answering style with that required but also to improve its document processing capabilities. Consequently, our approach outperforms all others. |
**Knowledge Reasoning.** The knowledge reasoning tasks rely on logical relations and evidence about factual knowledge to answer the given question. Existing work mainly uses specific datasets to evaluate the reasoning capacity of the corresponding type of knowledge, _e.g.,_ CSQA [504]/StrategyQA [185] for commonsense knowledge reasoning and ScienceQA [565] for science knowledge reasoning. In addition to the accuracy of the predicted results, existing work [565] has also evaluated the quality of the generated reasoning process, via automatic metrics (_e.g., BLEU) or human evaluation. Typically, these tasks require LLMs to perform step-by-step reasoning based on factual knowledge, until reaching the answer to the given question. To elicit the step-by-step reasoning ability, chain-of-thought (CoT) prompting strategy [33] has been proposed for enhancing the complex reasoning capacity of LLMs. As discussed in Section 6.3, CoT involves the intermediate reasoning steps, which can be manually created [33] or automatically generated [674], into the prompts to guide LLMs to perform multi-step reasoning. Such a way largely improves the reasoning performance of LLMs, leading to new state-of-the-art results on several complex knowledge reasoning tasks [33, 56, 526]. Further, after reformulating knowledge reasoning tasks into code generation tasks, researchers have found that the performance of LLMs can be further improved [211], especially with the LLMs pre-trained on code. However, due to the complexity of knowledge reasoning tasks, the performance of current LLMs still lags behind human results on tasks such as commonsense reasoning [675, 56, 33]. As a common type of mistakes, LLMs might generate inaccurate intermediate steps, leading to a wrong final result. To address this issue, existing work has proposed special decoding or ensemble strategies to improve the accuracy of the whole reasoning chain [436, 437]. |
We trained and evaluated small Infini-Transformer models on PG19 (Rae et al., 2019) and Arxiv-math (Wu et al., 2022) benchmarks. Our setup closely resembles that of Memorizing Transformers (Wu et al., 2022). Namely, all our models have 12 layers and 8 attention heads of dimension 128 each and FFNs with hidden layer 4096. |
where \(q\) is the quantization position in the current iteration while \(F\) denotes the remaining non-quantized weights. \(H_{F}\) is the hessian matrix of the objective function. |
where the actual answer \(y_{k+1}\) is left as a blank to be predicted by the LLM. Since the performance of ICL heavily relies on demonstrations, it is important to properly design them in the prompts. According to the construction process in Equation (12), we focus on three major aspects of formatting demonstrations in the prompts, including how to select examples that make up demonstrations, format each example into the prompt with the function \(f(\cdot)\), and arrange demonstrations in a reasonable order. |
* Evol-Instruct (Xu et al., 2023) includes 218 real-world human instructions from diverse sources such as online open-source projects, platforms, and forums. * Vicuna (Chiang et al., 2023) includes 80 diverse instructions generated by GPT-4 through prompt engineering. * Self-Instruct (Wang et al., 2022) includes 252 expert-written instructions motivated by user-oriented applications. * Koala (Geng et al., 2023) includes 180 conversation-style real user instructions that were posted online. |
Data Optimal Regime.Unlike prior works that train language models in either "compute optimal regime" [HBM\({}^{+}\)22] or "over-train regime", we mainly focus on the quality of data for a _given scale_.2 We try to calibrate the training data to be closer to the "data optimal" regime for small models. In particular, we filter the web data to contain the correct level of "knowledge" and keep more web pages that could potentially improve the "reasoning ability" for the model. As an example, the result of a game in premier league in a particular day might be good training data for frontier models, but we need to remove such information to leave more model capacity for "reasoning" for the mini size models. We compare our approach with Llama-2 in Figure 2. |
We begin with a random sample of 1000 articles from Huggingface's Wikipedia dataset (20220301.en, (Foundation)). We use GPT-4 to generate questions related to each field (dates, names, and cities) and filter out responses where the excerpt is not exactly found in the context. To reduce ambiguity when matching groundtruth answers, we restrict the answers to fit certain formats. For dates, we require that the answer adheres to a four-digit year (YYYY). For names, we require a first and last name (eg. George Washington). For cities, we remove any other identities (eg. Seattle, not Seattle, WA). For each domain, among the remaining question-answer pairs that fit these criteria, we randomly sample 200 for our evaluation set. |
While generic models Touvron et al. (2023); Taylor et al. (2022); Zhang et al. (2022); Gu and Dao (2023); Almazrouei et al. (2023); Zhang et al. (2024); Faysse et al. (2024) gain some exposure to legal data during their training, it typically only represents a minor fraction of the overall data. A straightforward method to enhance performance for legal tasks is to perform additional training focusing on legal data. This approach, particularly focused on decoder models, has been successfully used in various fields such as medicine Chen et al. (2023); Ji et al. (2023), translation Xu et al. (2023); Wu et al. (2024), and coding Roziere et al. (2023). The key advantage of this approach is its scalability and independence from the specific characteristics of the training data. Other research on domain adaptation has attempted to specialize language models via pretext tasks. However, these efforts often rely on smaller-scale approaches Niklaus and Giofre (2023), are computationally expensive Vu et al. (2020); Lu et al. (2023), or lack scalability Cheng et al. (2023); Cui et al. (2023); Nishida et al. (2019). |
With this in mind, it is likely that future versions of the AI Safety Benchmark will not be fully open. We are considering various strategies for how to manage their release in order to protect the benchmark's integrity while being as transparent as possible. Options under consideration include withholding a random portion of the benchmark dataset; withholdingcertain types of prompts; delaying public release of the prompts for a set period of time (e.g., six months); and imposing dataset deprecation conditions if the benchmark's integrity becomes questionable. Finally, starting from v0.5 we are establishing a set of requirements that anyone using the AI Safety Benchmark must adhere to. |
As shown Figure 1, our Infini-attention computes both local and global context states and combine them for its output. Similar to multi-head attention (MHA), it maintains \(H\) number of parallel compressive memory per attention layer (\(H\) is the number of attention heads) in addition to the dot-product attention. |
**Baseline Performance.** As shown in Table 2, we find that \(M^{L}\) (equivalent to FLAN-T5 XXL) does not benefit from the original CoT prompting, as its EM and F1 scores dropped in both datasets (except for answer inclusion score for HotpotQA) when compared to standard prompting. This is consistent with previous research findings that models with less than 50B parameters exhibit limited reasoning capabilities. We also observe that the performance drop is more significant with 2WikiMultihopQA (nearly 10% for EM and F1) than HotpotQA. Based on our manual inspection of incorrect predictions, \(M^{L}\) was prone to repeat sentences in the context and fail to provide a final answer to the questions. |
Immediate hazardsImmediate hazards are sources of harm that are already being presented by existing frontier and production-ready models. This includes enabling scams and fraud [86], terrorist activity [87, 88], disinformation campaigns [89, 90, 91], creation of child sexual abuse material[92], encouraging suicide and self-harm [93], cyber attacks and malware[94, 95], amongst many others [96]. Another concern is factual errors and "hallucinations". This is a substantial risk when models are faced with questions about events that happened after their training cutoff date if they do not have access to external sources of up-to-date information [6, 97, 98]. Generative AI has been shown to increase the scale and severity of these hazards by reducing organizational and material barriers. For instance, the media has reported that criminals have used text-to-speech models to run realistic banking scans where they mass-call people and pretend to be one of their relations in need of immediate financial assistance [99]. The risk of bias, unfairness, and discrimination in AI models is a longstanding concern, supported by a large body of research [100, 101, 44, 102]. Recent work also shows that out-the-box models can also be easily adjusted with a small fine-tuning budget to readily generate toxic, hateful, offensive, and deeply biased content [103, 104, 68]. And substantial work has focused on developing human- and machine-understandable attack methods to cause models to regurgitate private information [105], 'forget' their safety filters [106] or reveal vulnerabilities in their design [107]. |
By sharing artifacts from all pipeline stages, we aim to encourage open research and reduce duplicated, often costly efforts, by academics and practitioners. |
* Harmful and unpredictable content: Though our model can reject unethical queries, these models may still generate harmful or misaligned with human preferences and values. This issue may arise from biases in the training data or the models' inability to discern appropriate outputs in certain contexts. * Insufficient training: Due to constraints in computing power and data availability, the training of the models may not be sufficient for optimal performance. As a result, there is still room for improvement in the Chinese understanding capabilities of the models. * Lack of robustness: The models may exhibit brittleness in some situations, producing inconsistent or nonsensical outputs when faced with adversarial inputs or rare language phenomena. * Comprehensive evaluation: Evaluating large language models is an important topic in the current era. While we have seen many evaluation benchmarks for LLMs, their comprehensiveness and appropriateness for LLMs should be well-studied and examined. A more diverse and comprehensive LLM evaluation dataset and benchmark will have a great positive effect on shaping the future of LLM research. * Scalability and efficiency: Although we applied LoRA and quantization to make the model more accessible to a broader community, when combined with the original LLaMA, the models' large size and complexity can lead to difficulties in deployment, especially for users with limited computational resources. This issue may hinder the accessibility and widespread adoption of the models in various applications. |
A: I understand there is the ones day one 3 plus 4 eats who makes and so on and so on. |
In cases where the background description of a question contains links to webpages, our system scrapes them, parses the clean texts, and presents the summaries to the reasoning model. We take measures to ensure that this leaks no information beyond the retrieval range. First, we maintain a whitelist of news websites that publish timestamped articles and only retrieve from the whitelist. Second, our system checks the publish date of each article and discard it if the date is not available or outside the retrieval range. |
Full results are summarised in Table 10. One may notice that our reported forward task accuracies are slightly higher than those reported in Allen-Zhu & Li (2023a). This improvement is attributed to our use of a larger batch size, smaller weight decay, and the best result among three runs. |
In this part, we present discussions regarding two fundamental questions related to CoT prompting, _i.e._, _"when does CoT prompting work for LLMs"_ and _"why can LLMs perform CoT reasoning"_. |
fewer steps to converge. On the contrary, our method demands additional steps to reach a higher convergence value. This may be attributed to the exploration through masking, enabling us to gather new information from a limited dataset continually. Conversely, extending the training epochs for SFT introduces variability without yielding superior results. In our implementation details, we allow MFT to train for an extended number of epochs to ensure convergence of results as detailed in Tab. 8. |
_Remark 3.5_ (why "random data").: Studying a lower bound for a fixed dataset \(\mathcal{Z}\) is impossible -- a model could hard-code \(\mathcal{Z}\) into its architecture even without any trainable parameter. Therefore, it is necessary to consider a lower bound with respect to a _distribution_ over datasets. |
**Weaviate**[242] is an open-source, GraphQL-based vector search engine that enables similarity search on high-dimensional data. While it is open-source, the commercial version offers additional features, support, and managed services. |
We propose sDPO where we use the preference data in a stepwise way instead of all at once. We show that applying sDPO results in more performant models than DPO in terms of H4 score. We also empirically exhibit that sDPO results in more aligned reference models by comparing mean \(\gamma_{\pi_{ref}}\). |
**Comparing with Advanced Translation Models** Our model, ALMA-13B-R, is benchmarked against the most advanced current models, demonstrating performance comparable to GPT-4 and WMT winners. It surpasses leading commercial translation tools such as Google Translate in many cases and top multilingual translation models like NLLB, MADLAD-10B and GPT-3.5. |
In the rest of this paper, Section II presents an overview of state of the art of LLMs, focusing on three LLM families (GPT, LLaMA and PaLM) and other representative models. Section III discusses how LLMs are built. Section IV discusses how LLMs are used, and augmented for real-world applications Sections V and VI review popular datasets and benchmarks for evaluating LLMs, and summarize the reported LLM evaluation results. Finally, Section VII concludes the paper by summarizing the challenges and future research directions. |
Figure 1: **Overview of our retrieval and reasoning systems**. Our retrieval system retrieves summarized new articles and feeds them into the reasoning system, which prompts LMs for reasonings and predictions that are aggregated into a final forecast. |
As the performance is related to the prior probability, decomposing low-probability tasks into multiple high-probability subtasks and addressing them hierarchically can boost performance. Moreover, subtasks with varying prior probabilities can be addressed using different models. Additionally, cost savings can be achieved by using simpler, less expensive models for the easier, higher-probability subtasks. |
- He might not want to share the stage with candidates he perceives as less significant, which could dilute his brand or message. |
The focus of this paper is on the datasets used to train language models. While recent work focused on documenting the potential harms that could arise from problematic datasets (Bender and Friedman, 2018; Gebru et al., 2020), less work has been done to quantitatively analyze properties of real language modelling datasets, like Dodge et al. (2021) has done for C4. Our paper provides analysis on one particular axis, that of data duplication. |
[System] We would like you to provide feedback on the Assistant's performance regarding the above question and standard answer. Please rate the Assistant on helpfulness, relevance, accuracy, level of detail, and the use of { target language }. The Assistant will receive an overall score on a scale of 1 to 10, with higher scores indicating better overall performance. Please first output a single line containing only a value representing the Assistant's score. In the following line, please provide a comprehensive explanation of your evaluation, avoiding any potential biases and ensuring that the order of answers does not affect your judgment. |
To study what types of instructions are generated and how diverse they are, we identify the verb-noun structure in the generated instructions. We use the Berkeley Neural Parser7Kitaev and Klein (2018); Kitaev et al. (2019) to parse the instructions and then extract the verb that is closest to the root as well as its first direct noun object. 26,559 out of the 52,445 instructions contain such structure; other instructions usually contain more complex clauses (e.g., "Classify whether this tweet contains political content or not.") or are framed as questions (e.g., "Which of these statements are true?"). We plot the top 20 most common root verbs and their top 4 direct noun objects in Figure 3, which account for 14% of the entire set. Overall, we see quite diverse intents and textual formats in these instructions. |
**Results with LLaMA-based Target LLMs.** Table 1 summarizes the performance of CodecLM and the comparing baselines with 2000 synthetic data for instruction tuning. All methods are trained on LLaMA-7B or -13B as the target LLM and compared against Gemini-Pro, the strong LLM that generates the data. CodecLM outperforms comparing methods consistently on all benchmarks, with two target LLMs of different sizes. The consistently superior performance of CodecLM highlights its generalizability to different downstream instruction distributions and target LLMs. Both Tree-Instruct and variants of WizardLM focus on the importance of instruction complexity, however, their performances are not always better than Alpagasus with simple instructions, especially with larger target LLM. This observation indicates that the effectiveness of data cannot be solely determined by instruction complexity, and validates the motivation of our design of Self-Rubrics and Contrastive Filtering. Moreover, the win of WizardLM+ over WizardLM confirms the efficacy of instruction distribution matching via instruction metadata. When shifting the target LLM from LLaMA-7B to -13B, all methods get a significant performance boost, which accords with prior discoveries on scaling model size Wei et al. (2021). |
In this case, knowing the geography of Hawaii, as well as immediate effects of wildfires, enables a model to give the first two options a lower likelihood. This process of elimination increases the probability of choosing one of the remaining options (the third option is the correct answer). |
We evaluate the models on financial tasks to evaluate the effectiveness of our domain-adaptive continual pre-training. We adopt the _FLARE_ framework [33] to evaluate our models. FLARE extends the LLM evaluation framework _lm-evaluation-harmess3_ by including various financial tasks. We follow their instruction prompt, data split, and metric computation for comparison. We consider following 4 tasks used in [31; 33]: (1) **Financial Phrase Bank**. FPB is a sentiment classification task on financial news [19]. The sentiment reflects whether the news is considered as positive/neutral/negative by investors. (2) **FiQA SA.** An aspect based sentiment classification task based on financial news and headlines [18]. (3) **Headline.** Binary classification task on whether a headline on a financial entity contains certain information [25]. Each news article is associated with 9 tags like "price or not", "price up", "price down", "price stable", "past price", and "asset". (4) **NER.** Financial named entity extraction task is based on credit risk assessment section of SEC reports. Words in this task are annotated with PER, LOC, ORG, and MISC. |
Open domain instruction fine-tuneOur work belongs to this research line. OpenAI has hired many annotators and written many instructions with corresponding correct responses. These human-created instructions have diverse forms and rich task types. Based on this dataset, OpenAI trained GPT-3 [1] into InstructGPT [2], which can process a variety of real user instructions and led to the success of ChatGPT. Since these outstanding works from OpenAI were not open-sourced, Alpaca [31] and Vicuna [22] subsequently actively explored open-domain instruction fine-tuning based on the open-source LLM LLaMA [4]. Alpaca used a dataset of 50k instructions generated from a limited (e.g., 175 samples) seed set of manually-written instructions. Vicuna used 70k user-shared conversations with ChatGPT collected from ShareGPT.com. Our work is different from InstructGPT and Vicuna in that we use AI-generated data for instruction fine-tuning. Unlike Alpaca's self-instruct [32] generation method, _Evol-Instruct_ can control the difficulty and complexity level of the generated instructions. |
* First release on March 31, 2023: the initial version. * Update on April 9, 2023: add the affiliation information, revise Figure 3 and Table 1 and clarify the corresponding selection criterion for LLMs, improve the writing, and correct some minor errors. * Update on April 11, 2023: correct the errors for library resources. * Update on April 12, 2023: revise Figure 3 and Table 1, and clarify the release date of LLMs.
* Update on April 16, 2023: add a new Section 2.2 about the technical evolution of GPT-series models. * Update on April 24, 2023: add the discussion about scaling laws and add some explanations about the model sizes for emergent abilities (Section 2.1); add an illustrative figure for the attention patterns for different architectures in Figure 9, and add the detailed formulas in Table 6. * Update on April 25, 2023: revise some copy errors in figures and tables. * Update on April 27, 2023: add efficient tuning in Section 5.3.
* Update on April 28, 2023: revise Section 5.3.
* Update on May 7, 2023: revise Table 1, Table 2, and some minor points. * Update on June 29, 2023 (major revision):
* Section 1: add Figure 1 for the trends of published LLM papers in arXiv;
* Section 2: add Figure 4 for GPT's evolution and the corresponding discussion;
* Section 3: add Figure 5 for LLMA family and the corresponding discussion;
* Section 5: add latest discussion about the synthetic data formatting of instruction tuning in Section 5.1.1, the empirical analysis for instruction tuning in Section 5.1.4, parameter-efficient model adaptation in Section 5.3 and memory-efficient adaptation in Section 5.4;
* Section 6: add latest discussion about the underlying mechanism of LCL 6.2.3, planning for complex task solving in Section 6.4;
* Section 7: update Table 14 for representative datasets for evaluating advanced abilities of LLMs, and empirical ability evaluation in Section 7.4;
* Section 6.1.1: add prompt design;
* Section 8: add the discussions on applications of LLMs in finance and scientific research domains;
* Update on September 10, 2023 (major revision):
* Claim the copyrights of the figures and tables in this paper. * Add latest LLMs, techniques and their descriptions in Section 3, Section 4, Section 5, Section 6 and Section 7;
* Section 4: add latest discussion about the decoding strategy in Section 4.2.5;
* Section 5: add latest discussion about the practical tricks for instruction tuning in Section 5.1.2, the empirical analysis on LLMA (13B) for instruction tuning in Section 5.1.4, practical strategies for RLHF in Section 5.2.3, alignment without RLHF in Section 5.2.4 and remarks on SFT and RLHF in Section 5.2.5;
* Section 6: update the content about the planning for complex task solving in Section 6.4;
* Section 7: add discussions about evaluation approaches in Section 7.3.2, Table 15 for the category of existing evaluation work, and update empirical ability evaluation in Section 7.4 and the results on Table 16;
* Section 6.1.1: add new prompt examples in Table 12;
* Update on November 23, 2023 (this version):
* Section 1: add Figure 2 for the evolution process of four generations of language models;
* Section 2: add more discussion about scaling laws and how emergent abilities relate to scaling laws;
* Section 3: add latest LLMs in Figure 3 and Table 1, latest APIs in Section 3.1, commonly used datasets
for instruction tuning and alignment tuning in Section 3.3, and several libraries in Section 3.4;
* Section 4: add latest discussion about the data scheduling, including data mixtures and data curriculum in Section 4.1.3; add summary of data preparation in Section 4.1.4; add discussion about modeling long context in Section 4.2.4; add discussion about decoding efficiency issues and add latest decoding strategies in Section 4.2.5;
* Section 5: add latest discussion about instance construction and tuning strategies in Section 5.1; add latest discussion about process-supervised RLHF in Section 5.2.3, and the empirical study on quantized LLaMA models (7B and 13B) in Section 5.4.3;
* Section 6: add latest discussion about prompt optimization in Section 6.1.2, and update the content about chain-of-thought prompting in Section 6.3;
* Section 8: add latest discussion about LLM for research directions in Section 8.1;
* Section 9: revise the content in the several aspects. |
Figure 2 presents an ablation study investigating the impact of the number of data evolution rounds. The first round of evolved data contains 38k samples. The second round contains 58k. The third round contains 78k. The fourth round contains 98k. For consistency, all models undergo fine-tuning with 200 steps. The results reveal that the highest pass@1 score on humaneval is achieved after three rounds of data evolution. Based on this observation, we select the data that evolved during the third round as the ultimate dataset. |
For all experiments, we used a Byte Pair Encoding (BPE) vocabulary trained on C4-NearDup with a budget of 50K tokens, which resulted in a vocabulary the same size as GPT-2's. We trained with a maximum sequence length of 512 tokens (for longer documents, we randomly extracted subsequences of this length.) Further training details can be found in Appendix C. |
Figure 1: Scores of systems under test (SUTs) across the different hazard categories)within the AI Safety Benchmark v0.5. The bottom row reports the overall benchmark performance (i.e., lowest grade across hazard categories). Reference models for scoring are denoted with a (\(*\)). |
**(RQ1)** What is the effectiveness of RAG and fine-tuning with synthetic data on QA for low-frequency factual knowledge? |
**In-depth Surpassing Human Instructions.** To study the depth of the instruction evolving process, we use ChatGPT to help us judge the difficulty and complexity level of each instruction. The used prompt please refer to Appendix E. The Figure 6(a) and 6(b) illustrate that _Evol-Instruct_ generated instructions that were more complex than those created by human participants in ShareGPT. Moreover, the depth of the instructions increase significantly with each iteration of the evolution process. |
* We present the first systematic study on the scaling property of raw agents instantiated by LLMs. We find that the performance scales with the increase of agents, using the simple(st) way of sampling and voting. * We explore the compatibility of our method with existing complicated methods that stimulate the potential of LLMs, revealing that our method can enhance these methods to achieve further performance improvements. * We analyze the effectiveness of our method in tackling problems at varying difficulties and then distill the properties behind, based upon which, we propose further optimization methods that can facilitate the occurrence of our finding. |
Instance Generation.Given the instructions and their task type, we generate instances for each instruction independently. This is challenging because it requires the model to understand what the target task is, based on the instruction, figure out what additional input fields are needed and generate them, and finally complete the task by producing the output. We found that pretrained LMs can achieve this to a large extent when prompted with instruction-input-output in-context examples from other tasks. A natural way to do this is the **Input-first Approach**, where we can ask an LM to come up with the input fields first based on the instruction, and then produce the corresponding output. This generation order is similar to how models are used to respond to instruction and input, but here with in-context examples from other tasks. The prompting template is shown in Table 7. |
Subsets and Splits