Datasets:
Translation-EnKo
/
arxiv-translation

Dataset card Data Studio Files Community
2
text
stringlengths
4
4.47k
contrast to these works, Hu et al. (2024) proposed a method called Gradient Guided Prompt Perturbation (GGPP), a way to perturb the RALM, which was experimentally found to be effective in improving the situation by utilizing the SAT probe (Yuksekgonul et al., 2023) and activation (ACT) classifier. A method is proposed to detect this perturbation by prompting the internal state of the perturbed RALM. In addition, by proposing and improving the evaluation method of RALM and the related baseline can also help improve the robustness of the model, Chen et al. (2024) made a series of evaluation system for RALM by focusing on the robustness.
For _fine-tuning_ of EEVE-Korean models, we employed the Direct Preference Optimization (DPO; Rafailov et al.2023) based on LLaM-Factory implementation. To further enhance the models' capabilities of following Korean instructions, we translated the publicly available instruction datasets, specifically Orca5(Mukherjee et al., 2023; Lian et al., 2023) and UltraFeedback6(Cui et al., 2023) into Korean. In the process of translating these datasets into Korean, ensuring the integrity of programming code formats and correcting translation errors, such as instances where both the source and target languages were inadvertently translated into Korean, was crucial for maintaining the quality and effectiveness of our fine-tuned models. We named the fine-tuned models as EEVE-Korean-Instruct.
\(\bullet\)_LLM based interactive approaches._ Most existing approaches train LLMs in isolation, where LLMs are not present in actual environments to improve themselves through external feedback signals. As a comparison, humans learn social norms and values from interactions with others in social environments [387]. To mimic such a learning approach, Stable Alignment [179] builds a simulated interaction environment consisting of a number of LLM agents, where AI agents keep interacting with and each other, receiving feedback on improvement. Once a central agent receives an instruction, it produces a response and shares it with nearby agents. These critic agents generate feedback comprising ratings about the response and revision suggestions. Then the central agent would revise the original response following these suggestions. Such an alignment approach can be also extended to real-world environment with humans.
Figure 1: A schematic of generating modified documents for each dataset. A question is posed to the LLM with and without a reference document containing information relevant to the query. This document is then perturbed to contain modified information and given as context to the LLM. We then observe whether the LLM prefers the modified information or its own prior answer.
Note that the first category of dataset (NQ, Trivia QA, and HotpotQA) is a relatively general domain whereas the latter two domains are on very domain specific documents.
We perform experiments on two mathematical reasoning benchmarks, namely GSM8k [42] and MATH [43], the results demonstrate that our _WizardMath_ outperforms all other open-source LLMs, achieving state-of-the-art performance. Specifically, _WizardMath_ observe a substantial improvement in pass@1 with an increase of +24.8 (81.6. vs. 56.8) on GSM8k, and +9.2 (22.7 vs. 13.5) on MATH. Notably, our model even also significantly surpasses OpenAI's ChatGPT-3.5 [5], Anthropic's Claude Instant-1 [39], and Google's PaLM-2 [44] in terms of pass@1 on GSM8k.
(2) **Training scales required for effective transfer.** We find that further Chinese pretraining with 100 billion tokens or fewer is insufficient to significantly improve LLaMA's knowledge level. However, enhancing LLaMA's response quality (i.e., language generation capability), requires only hundreds of thousands of instruction data rather than a large-scale further pretraining.
The results of the experiments are shown in Table 1. We can see that, despite the A-2 and A-1 have undergone the same SFT distribution, adding SFT data to the decay stage pushes the boundary. Comparison between B-2 and B-3 demonstrate that the deficiency of only SFT is not due to the insufficient training tokens in SFT stage.
**Code-based Approaches.** Although text-based approaches sound intuitive, they cannot guarantee faithful execution of the plan, which may lead to failure even when the plan is sound. To address this issue, code-based approaches have been proposed to generate more verifiable plans in the form of executable code in programming languages, _e.g.,_ Python or PDDL. In this way, LLMs are first prompted to generate the program and then utilize a deterministic solver to execute it. For example, FaithVolCt [442] and PAL [443] decompose a reasoning task into two stages: at the first stage, the LLM generates a plan conditioned on the query; at the second stage, a deterministic solver executes the plan to derive the final answer. Furthermore, code-based approaches can be applied to embodied agents in a similar way. For example, PROGPROMPT [530] and LLM+P [531] first utilize LLMs to generate plans in the form of python functions or PDDL files, and then leverage a virtual agent or classical planner to solve the problem according to the code-based plans.
In Figure 13, we display the dynamic changes in token selection tendencies throughout the SLM training process. We chose four checkpoints during the training process (0%, 33%, 66%, and 100%) to analyze the current tendencies in token selection. The preferences for token selection are indicated by different colors, ranging from high to low preference, typically represented as deep blue, blue, black, orange, and dark orange, respectively.
To address the above mentioned challenges, we propose AIOS, an LLM agent operating system (Figure 2) to provide module isolation and aggregations of LLM and OS functionalities. To address the potential conflicts arising between tasks associated with LLM and those unrelated to LLM, we propose the design of an LLM-specific kernel. This kernel segregates the OS-like duties, particularly those related to the oversight of LLM agents, their corresponding resources, and development tool-kits. Through this segregation, the LLM kernel aims to enhance the management and coordination of LLM-related activities. Within the proposed LLM kernel, we have devised a suite of modules, each dedicated to addressing the distinct functions pertinent to LLM operations. An overview of these modules and their respective functionalities is presented in the following, with the comprehensive discussion of their underlying mechanisms detailed in Section 4.
Comparing our model's performance on both English and Chinese benchmarks with other models reveals a notably smaller gap in performance across multi-disciplinary datasets such as MMLU and CMMLU, as shown in Table 2. While other models exhibit significant disparities, particularly in language understanding and reasoning benchmarks, our model maintains a consistent performance, suggesting a balanced capability across diverse domains. This contrasts with other models that show pronounced variability, such as in the HellaSwag dataset, where our model closely rivals or outperforms alternatives like MiniCPM (min, 2024) and Phi-2, showcasing superior or competitive reasoning abilities. Similarly, in domain-specific evaluations (C-Eval and CMMLU), our model demonstrates commonable performance, outpacing models like TinyLlama-1.1B and Bloom-1.7B in comprehending and generating content that requires a nuanced understanding of cultural and domain-specific contexts. This balanced proficiency underscores the model's versatility and adaptability, positioning it as a strong contender in the landscape of AI language models, with a capacity for both broad applicability and deep, domain-specific knowledge.
XLMs [47] extended BERT to cross-lingual language models using two methods: (1) a unsupervised method that only relies on monolingual data, and (2) a supervised method that leverages parallel data with a new cross-lingual language model objective, as illustrated in Fig 5. XLMs had obtained state-of-the-art results on cross-lingual classification, unsupervised and supervised machine translation, at the time they were proposed.
As BloombergGPT is evaluated using an in-house data split, and the calculation details of reported metrics may not be identical, direct comparisons of their results with ours are not feasible. To adequately assess the efficacy of continual pre-training, we benchmark BloombergGPT's performance against the FLARE framework. This involves evaluating OPT-66B and GPT-NeoX-20B's performance, as obtained from FLARE, and comparing it to the results reported in [31]. This rigorous benchmarking ensures a fair and comprehensive evaluation, providing valuable insights into the effectiveness of our continual pre-training approach in relation to financial LLMs trained from scratch.
cluding GPT-4 (Achiam et al., 2023), were trained on predominantly human-generated text; similarly, diffusion models like DALL-E (Ramesh et al., 2021), Stable Diffusion (Rombach et al., 2022), Midjourney (Midjourney, 2023) are trained on web-scale image datasets. These training corpora already potentially exhaust all the available clean data on the internet. A growing number of synthetic data generated with these increasingly popular models starts to populate the web, often indistinguishable from "real" data. We have thus already raced into the future where our training corpora are irreversibly mixed with synthetic data and this situation stands to get worse. Recent works call attention to the potential dramatic deterioration in the resulting models, an effect referred to as _"model collapse"_(Shumailov et al., 2023). Facets of this phenomenon have been demonstrated _empirically_ in various settings (Hataya et al., 2023; Martinez et al., 2023; Bohacek and Farid, 2023; Briesch et al., 2023; Guo et al., 2023). Theoretically, a few works are emerging to analyze the effect of iterative training on self-generated (or mixed) data (see Related Work): (Shumailov et al., 2023) coin the term _"model collapse"_ to characterize complete reversion to the mean, Alemohammad et al. (2023) analyze _"self-consuming loops"_ and Bertrand et al. (2023) show that iterative synthetic training leads to a _"clueless generator"_.
1. The first prompt containing the question is sent to the model (step 1 of Figure 1). 2. The Adapt-LLM evaluates the prompt to determine whether additional context is necessary to answer the question effectively (step 2). 3. If the model determines that context is not required, it directly produces a response to the question by leveraging its parametric memory (step 3). 4. If context is deemed necessary, the Adapt-LLM model returns a special token, represented as \(\langle\text{RET}\rangle\), and an off-the-shelf IR system is used to retrieve pertinent context based on the question (step 4); the context is then combined with the original question prompt to form a comprehensive representation for answer generation (step 5).
(1) **The impact of vocabulary extension on transfer.** We find that further pretraining with 0.5 billion Chinese tokens on the original vocabulary significantly outperforms performance on the extended vocabulary, even though the latter has been further pretrained on over 30 billion tokens. This suggests that vocabulary extension might not be a suitable choice for small-scale incremental pretraining in the order of tens of billions.
_Smarter parameterizations for inversion._ Future work might consider exploiting the fact that inputting a single suffix into a LM outputs multiple next-token predictions, one at each position, not just at the end. Additional research may find that utilizing a parameterization that integrates token embeddings with probability values, so that the inversion model 'knows' which value corresponds which word, could be useful.
However, unlike the first-hop reasoning (SS5), the second-hop reasoning does not strengthen with increasing model size; when scaling from 7B to 13Band 70B, the maximum relative frequency remains relatively stable at 0.64 (7B), 0.65 (13B), and 0.61 (70B), as shown in Figure 2(c). It is worth noting that this finding aligns with the observation of Ofir Press et al. (2023), that the single-hop question answering performance improves faster than the multi-hop performance as the model size increases, and thus the _compositionality gap_ (the ratio of how often models can correctly answer all sub-problems but not generate the overall solution) does not decrease with increasing model size.
The amount of knowledge elicited through the standard instruction-tuning is still limited, even though the perplexity of documents is minimized, a phenomenon we refer to as the "perplexity curse". Our next question is how can we improve the ability of LLMs to absorb knowledge from documents to mitigate the perplexity curse. The main challenge is the gap between the way knowledge is presented in raw documents and how it is accessed through question-answering. We found that QA pairs are generally straightforward, while documents tend to be more complex and cluttered, weaving many factual statements together in a more intricate manner. Using Fig. 3 as an example, the answer to the question "who handled the editing of Oppenheimer" is included in a sentence in the middle of the article "Editing was handled by Jennifer Lame...", which does not explicitly mention "Oppenheimer". During training, LLMs must understand the context and deduce that "editing" refers to "the editing of Oppenheimer" to effectively encode this knowledge in the parameters.
While human evaluation can generally offer reliable quality assessment, it is also often hindered by high annotation costs, significant time requirements, and annotation inconsistencies [887]. In contrast, automatic evaluation can be employed as a scalable alternative to human evaluation. Traditional automatic evaluations have relied on reference-based metrics (_e.g._, BLEU and ROUGE). Recently, with the emergence of LLMs as general task solvers highlights their potential as automatic evaluators [647, 727], making it promising to conduct LLM based evaluation. In the following part, we will introduce the recent progress on LLM for evaluation, including evaluation formats, methods, meta-evaluation, and the remaining issues.
The idea is that given an auxiliary knowledge base and an input query, we use the RAG architecture to find documents within the knowledge base that resemble the input query. These documents are then added to the input query, thus giving the model further context about the subject of the query.
**MODEL ANSWER:** Violetta needs to buy 5 * 2 = <<5*2=10>>10 crayons. She will get change of 20 - 10 = $*20-10=10>>10.
Footnote 1: Details about how to access Falcon LLM open source is available on falconllm.tii.ae
Focusing on building a large-scale high-quality web pretraining dataset, we extend upon the state-of-the-art in three ways: (1) we aggregate and combine best-practices for document preparation and filtering across multiple pipelines, and introduce line-wise corrections; (2) we combine both exact and fuzzy deduplication at very large-scale; (3) the scale of our final dataset is unique, with a total 5,000 billion tokens, and a 600 billion tokens extract available for public use with permissive licensing. Training large models on RefinedWeb also lead us to challenge the commonly held belief that web data is strictly worse than curated corpora.
The midpoint of a line segment with endpoints \((x_{1},y_{1})\) and \((x_{2},y_{2})\) is given by the formula:
where \(TV(a,b):=\sum_{j}|a_{j}-b_{j}|\) is the total-variation distance and the expectation is over the randomness in \(q_{T}\). An asset here is that (Berend and Kontorovich, 2012) can be used to control the quantities \(\operatorname{\mathbb{E}}\left[TV(q_{T}(\cdot\mid i),p(\cdot\mid i))\right]\). Note that TV is upper-bounded by the square-root of KL-divergence, thanks to _Pinker's inequality_. This gives indication that our results could also apply in the setting of autoregressive models with perplexity loss, like LLMs.
List of Domains used to Filter Financial NewsWe use the following keywords to identify subdomains and urls: economy, market, finance, money, wealth, invest, business, industry.
Figure 3: **Performance of base models on LegalBench-Instruct.
To this end, this technical report presents a novel approach for efficient and effective vocabulary expansion, namely EEVE, which can better train the embeddings of newly added tokens. For ease of adaptation, we utilize subword-based embedding initialization and design seven training stages with parameter freezing, which elaborately adjust the order and amount of parameters to be trained. We meticulously transfer the advanced capabilities of foundational models from English to Korean by initially focusing on the training of only input embeddings and progressively expanding to encompass the full parameters in the final stage.
1. Rephrased and Expanded Question: Is it likely that Kevin McCarthy, the former House Speaker, will make a public announcement or have his resignation reported by credible sources, indicating his departure from the United States House of Representatives, before the end of November 30, 2023? This question takes into account the current political climate, McCarthy's recent loss of the Speaker position, his consideration of not seeking re-election, and the potential impact of his decision on the Republican Party and the 20th District of California. The resolution of this market hinges on whether McCarthy will officially declare his intention to vacate his seat within the specified timeframe. 2.
The authors would like to thank many people who helped bring this work to fruition: Srini Iyer, Yuchen Zhang, Todor Mihaylov, Jacob Xu Moya Chen, Mansheej Paul, Mitchell Wortsman, Amro Abbas, Aaditya Singh, Myra Cheng, and Matthew Leavitt. The authors would also like to thank Surya Ganguli, Mona Diab, and Xian Li for initial brainstorming and are grateful for help with compute infrastructure given by Henry Estela and Victoria Lin. Lastly, the authors would like to thank anonymous reviewers for improving the quality and writing of this paper.
On the other end of the spectrum, it is worth mentioning that there are some attention-compatible architectural mechanisms that have been recently gaining steam and proving their value in creating better and more powerful LLMs. Probably the best example of such mechanism is Mixture of Experts (MoE). MoEs have been around in machine learning for years, even before the Deep Learning Era [213], but they have been gaining popularity since then, and particularly in the context of Transformer models and LLMs.
**Adapting Context Window.** Since Transformer-based LLMs have limited context windows, they can not directly integrate or utilize the entire information of the long sequences exceeding the context window. To alleviate the limitation, several methods adapting LLMs to long context have been proposed, as discussed below.
We choose \(11000\) clusters following previous work [1] and we note that this choice sticks to the heuristic that the number of clusters should roughly be the square root of the number of total points being clustered. We also note that in initial experiments for data selection at the 125M OPT model scale, we did not find a significant effect of number of clusters on the performance of our data selection methods (see Figure A1) this finding agrees with Abbas et al. [1] who notice significant overlap between datasets selected by SemDeDup with different number of clusters (see Figure A2 in Abbas et al. [1]).
where \(c=1-1/\beta\in(0,1)\) and \(\Gamma\) is the (upper) incomplete gamma function.
The Chain of Thought (CoT) technique, initially described in the paper "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models"[34] by Google researchers, represents a pivotal advancement in prompt engineering for Large Language Models (LLMs). This approach hinges on the understanding that LLMs, while proficient in token prediction, are not inherently designed for explicit reasoning. CoT addresses this by guiding the model through essential reasoning steps.
created and reorganized dynamically for each planning, execution, and evaluation phase [2, 9, 18]. This reorganization provides superior results because specialized agents are employed for certain tasks, and removed when they are no longer needed. By matching agents roles and skills to the task at hand, agent teams can achieve greater accuracy and decrease time to meet the goal. Key features of effective multi-agent architectures include clear leadership in agent teams, dynamic team construction, and effective information sharing between team members so that important information does not get lost in superfluous chatter.
In this section, we first review the popular architectures used for LLMs, and then discuss data and modeling techniques ranging from data preparation, tokenization, to pre-training, instruction tuning, and alignment.
**Model Architecture.** Due to the scalability and effectiveness, Transformer has become the de facto architecture for building LLMs. Various strategies have been proposed to improve the performance of this architecture, such as neural network configuration and scalable parallel training (see discussions in Section 4.2.2). However, Transformer still suffers from high training costs and slow inference rates. More efforts [251, 252] are still in need to develop improved model architectures for large-scale pre-training. Specially, system-level or hardware-level optimization (_e.g._, FlashAttention [284]) is worth more exploration to improve the efficiency of Transformer architectures. In addition, as an important basic capacity, existing LLMs typically maintain a long context window. For example, the most recent GPT-4 Turbo enables a long context of 128K tokens, and Claude 2.1 also supports the input up to 200K tokens. Although many efforts have been made to enhance the long context modeling ability of LLMs [264, 291], the resulting models still can't well process the information in the context window [299]. To address this issue, specific architecture adaptations or algorithms might be needed to enhance the modeling and utilization of long context information. Another worrying concern is that existing work mostly focuses on training LLMs with decoder-only Transformers. Despite the effectiveness, it severely limits the more wide, diverse explorations on alternative model architectures.
This publication is part of the project LESSEN with project number NWA.1389.20.183 of the research program NWA ORC 2020/21 which is (partly) financed by the Dutch Research Council (NWO).
One key benefit of the agent abstraction over prompting base language models is the agents' ability to solve complex problems by calling multiple tools. These tools enable the agent to interact with external data sources, send or retrieve information from existing APIs, and more. Problems that require extensive tool calling often go hand in hand with those that require complex reasoning.
AMIE, built upon the base LLM PaLM 2 [10], was instruction fine-tuned to enhance its capabilities for medical dialogue and reasoning. We refer to the PaLM-2 technical report for more details on the base LLM architecture.
In the logprob-free API, we have produced attacks capable of recovering logits and ultimately the embedding hidden dimension and embedding matrix up to a similarity transform. We now provide lower bounds on the minimum number of queries required by _any_ attacker attempting model stealing under the logprob-free API threat model.
Figure 8: **Perplexity on legal documents for pretrained backbones.** SaulLM-7B-Instruct outperforms other pretrained backbones on most types of legal documents, but is outperformed by Llama2-7B on Legislation. SaulLM-7B-Instruct exhibits a median perplexity of \(8.69\), having a reduction of \(5.5\) percent compared to Mistral-7B, \(9.20\), and \(10.8\) percent compared to Llama2-7B, with a median perplexity of \(9.74\).
Figure 1: **Continual pre-training decreases computational costs of updating the model while maintaining similar final validation and average evaluation performance. We report results for the Pile \(\cup\) SlimPajama(SP)/German(Ger.) baseline model trained on the union of both datasets which we consider to be an upper bound on performance. We also report performance for two continually pre-trained models. “PT on Pile” starts from a pre-trained Pile checkpoint and only uses learning rate re-warming and re-decaying, while “Replay (PT on Pile)” re-warms the learning rate, re-decays it, and uses 5% replay for SlimPajama and 25% replay for German. We observe that the combination of LR re-warming, re-decaying, and replay allows our continually pre-trained model to attain similar average performance to the baseline model while requiring substantially less compute.
LLM system call interface within the LLM kernel is designed to offer basic LLM call operation functions. This interface acts as a bridge between complex agent requests and the execution of different kernel's modules. As is shown in Table 2, analogous to OS system calls, LLM system calls offers a suite of basic functions that span across the kernel's modules, including agent management, context handling, memory and storage operations, and access control. The LLM system call list can be further expanded in the future to support more operations.
Colossal Cleaned Common Crawl (C4)is made up of 360M web documents, with an average length of 486 BPE tokens Raffel et al. (2020). C4 was introduced as a pre-training dataset for T5, a set of encoder-decoder models which have been widely used in fine-tuned downstream tasks. The dataset was previously deduplicated in a more sophisticated process than the prior two datasets. Each paragraph was hashed and paragraphs resulting in hash collisions were removed. This was followed by a pass that removed placeholder text, code, and prohibited words. See Dodge et al. (2021) for a detailed breakdown of the source text in C4.
Next, we need to find out the number of emeralds. We are told there are twice the number of emeralds than the rubies.
To assess SUTs' performance on each test, we first calculate the percentage of associated test items that are scored unsafe by the evaluator model. This results in a single percentage score for each test. It can be difficult to translate percentage scores into meaningful risk assessments, as this typically requires additional contextual information. For instance, a score of "5%" (i.e. 5% unsafe responses) on its own does not intrinsically indicate whether a SUT presents a low or high safety risk. To make the scores more interpretable we construct a grade for how SUTs perform on each test, on a five-point scale from "Low risk" to "High risk". How the grades should be calculated from the percentage scores is an open problem - and one that the WG devoted substantial time to discussing. We surfaced dozens of proposals, which can be broadly split into the three options presented below. We opted for a combination of the first option (i.e., using an absolute cutoff) and the third option (i.e., using a **reference model**) to grade SUTs' performance on each test.
We believe evolutionary algorithms will be able to discover more effective model merging solutions, and thus provide a path for automating the creation of more capable models. As a step towards this direction, in this work, we show that evolution can be employed to discover novel and unintuitive ways to merge various models to produce new models with a new combined ability. In this work, we present a methodology that leverages evolutionary algorithms to facilitate the merging of foundation models. Our approach is distinguished by its ability to navigate both parameter space (weights) and the data flow space (inference path), proposing a framework that integrates these two dimensions.
In each case, we observe a trade-off between language model fidelity and inversion performance.
Figure 7: **Test accuracy for GPT-3-350M using different hyperparameters.** Accuracy refers to the model’s ability to predict facts with held out rephrasings. **Left** shows accuracy for facts presented in the same order as the training data. **Right** shows accuracy for facts presented in the reverse order.
This paper introduces a simple method designed to enhance the performance of Large Language Models (LLMs). While the proposed method aims to improve the efficacy of LLMs in various tasks, it is necessary to acknowledge the potential risks. LLMs can sometimes produce outputs that, while plausible, may be factually incorrect or nonsensical. Such hallucinations can lead to the misguidance of decision-making processes and the propagation of biases. These concerns are particularly acute in the context of critical decision-making scenarios, where the accuracy and reliability of information are paramount. The broader adoption of LLMs, without adequate safeguards against these risks, could exacerbate these issues. Therefore, it is crucial to continue developing mechanisms to mitigate the potential adverse effects of LLM hallucinations to ensure that the deployment of these powerful models is both responsible and beneficial.
* to int8 has a negligible impact on their capacity. * to int4 reduces their capacity by more than 2x.
role-user: Given the following conversation between PersonU and PersonA: PersonU: Hello, I would like to know what to do if I do not agree with any decision. PersonA: disagree with our decision about your monthly income adjustment amounts? PersonU: no. Where can I find my SHIP contact information? PersonA: You can find your local SHIP contact information in the back of your Medicare & You 2020 Handbook online. PersonU: and how do they calculate the adjustments? Instead of having this entire conversation, how can PersonU get what he or she is looking for using a single question? Respond with that question.
T5 [52] is a Text-to-Text Transfer Transformer (T5) model, where transfer learning is effectively exploited for NLP via an introduction of a unified framework in which all NLP tasks are cast as a text-to-text generation task. mT5 [53] is a multilingual variant of T5, which is pre-trained on a new Common Crawl-based dataset consisting of texts in 101 languages.
\(\bullet\)_All the comparison models perform not well on very difficult reasoning tasks._ On MATH and HotpotQA, all models (including ChatGPT) perform not well. The two tasks are very difficult to solve, requiring accurate understanding of complex mathematical knowledge and performing multi-hop reasoning across documents, respectively. Further, these models also have a relatively weak performance on machine translation task (WMT). A possible reason is that WMT also contains many evaluation examples in minor languages, which might not be well covered in the pre-training data of these LLMs.
**Results with PaLM-based Models.** Table 2 summarizes the results of CodecLM and the best performing baselines in LLaMA-based experiments. We generate 1000 synthetic data due to computation budget. Since text-bison is a proprietary model that has been aligned with various techniques including instruction tuning, we also include it as a baseline approach. Interestingly, text-bison obtains strong performance across different benchmarks. Both Alpagasus and WizardLM+ underperform text-bison, suggesting it is non-trivial to improve upon a well-tuned LLM continually. CodecLM, on the contrary, outperforms text-bison in most cases, thanks to our core designs that adaptively tailor high quality data pairs to improve the target LLM.
In this section, we apply the PLND method to selected languages and models in order to confirm the existence of language-specific neurons and investigate the relationships between languages.
While previous work on LLM-based autonomous agent has obtained many remarkable successes, this field is still at its initial stage, and there are several significant challenges that need to be addressed in its development. In the following, we present many representative challenges.
Training with LoRA is much faster, memory-efficient, and produces smaller model weights (a few hundred MBs), that are easier to store and share. One property of low-rank matrices is that they can be represented as the product of two smaller matrices. This realization leads to the hypothesis that this delta between fine-tuned weights and initial pre-trained weights can be represented as the matrix product of two much smaller matrices. By focusing on updating these two smaller matrices rather than the entire original weight matrix, computational efficiency can be substantially improved.
Methods enhanced by our methodTo examine the comparability of our method, we study the integration of various typical methods from two distinct categories with our method:
- The calculated probability may be too confident given the volatile nature of the Middle East and the unpredictability of military engagements.
We perform all of our training runs on a version of CommonCrawl pre-processed with a CCNet [54] pipeline identical to the one used by Touvron et al. [50]. We add an additional step of MinHash-based de-duplication (see more details in Section A.1). Applying this common step before our experiments guarantees that any effects observed in our experiments complement the currently prevalent approach of MinHash-based data de-duplication strategies. Throughout the rest of this work, we refer to this dataset as _CC-dedup_.
Large Language Models For Instruction Fine-Tuning.The initial endeavors in instruction-following training work primarily focused on enhancing the language model's capacity for generalization across diverse tasks. This often involves the process of fine-tuning across substantially available Natural Language Processing datasets, and evaluates on the different NLP tasks. T5 [86] undertake the earliest attempts to train a range of NLP tasks, including Question and Answer, Document Summarization, and Sentiment Classification, by employing a consistent prompt format across all the data. Subsequently, instruction fine-tuning work such as FLAN [87], ExT5 [88], T0 [89], UnifiedQA [90], ZeroPrompt [91], and FLAN-T5 [92] emerged to adapt for a large number of downstream tasks. To address the challenge of misalignment between model outputs and human requirements, OpenAI manually annotates the instruction library to construct a diverse range of tasks. Simultaneously, Reinforcement Learning from Human Feedback technology is employed, which facilitate the rapid development of LLMs such as InstructGPT [2], ChatGPT\({}^{\ast}\), GPT-4 [3]. To reduce manual involvement, self-instruct [93] improves instruction-following through self-generated instructions. Alpaca [22] used a dataset of 50k instructions generated from a limited (e.g., 175 samples) seed set of manually-written instructions. Vicuna [23] used 70k user-shared conversations with ChatGPT collected from ShareGPT.com. Meanwhile, WizardLM [24] introduces the evol-instruct approach, which seeks to refine the existing instruction data by enhancing both its complexity and diversity.
When running the full layer-stealing attack, we confirmed that our extracted weights are nearly identical to the actual weights, with error \(<7\cdot 10^{-4}\), up to an \(h\times h\) matrix product as discussed previously. Table 4 reports the RMS between our extracted weight matrix and the actual model weights, after "aligning" the two by an \(h\times h\) transform.
User: {Question 1} Assistant: {Answer 1}... User: {Latest Question} Assistant: We use the {Latest Answer} from Assistant as the supervision for model outputs.
* _INT8 weight quantization can often yield very good results on LLMs, while the performance of lower precision weight quantization depends on specific methods_[414, 416, 417, 421]. In most cases, INT8 weight quantization can be effectively applied to reduce the memory footprint without performance degradation. While for INT4 (or INT3) weight quantization, existing methods rely on specific strategies to reduce the performance degradation, _e.g.,_ layerwise method [415, 417], activation-aware scaling [416] and low-rank adapter tuning [419]. Interestingly, LLMs seem to be less sensitive to low-bit weight quantization than small-sized language models [421]. In practice, with the same memory cost, it is suggested to use a larger language model with a lower quantization precision rather than a smaller language model with a higher quantization precision. For example, a 4-bit 60GB LLM is demonstrated to have better performance than a 8-bit 30GB LLM [422]. Moreover, focusing on emergent capabilities, the study [423] finds that in-context learning, step-by-step reasoning, and instruction following all seem to be seldom affected with 4-bit weight quantization. This result suggests that INT4 quantization exhibits a favorable trade-off in terms of both total bits and performance of emergent abilities. * _Activations are more difficult to be quantized than weights_[413, 414, 421]. It has been found that large outliers would occur for Transformer language models having a size of 6.7B or above [413]. This issue has been one of the most fundamental difficulties to quantize LLMs. To overcome this issue, various methods, _e.g.,_ mixed-precision decomposition [413], fine-grained quantization [413, 425] and difficulty migration [414], can be applied to alleviate the influence of outlier values. Since large outliers mainly exist in the activations of LLMs, small language models are more resistant to activation quantization [421, 423]. In practice, high-quality INT8 activation quantization is still a difficult task, though several methods can attain satisfying results. Further, lower precision activation quantization has still not been successfully explored, even for QAT methods [420].
Table 3 shows the results. Our model trained on the mixing data (240K MATH synthetic data + 240K GSM8K synthetic data) ranked as the second, just behind the GPT-4 and much better than other models. Additionally, we plot the correlation between GSM8K and Hungarian national high-school exam scores in Appendix B. The results show that there is no significant benchmark overfitting in our model.
The checkpoint used for evaluating OLMo-7B is trained until 2.46T tokens on the Dolma (Soldaini et al., 2024) dataset with a linear learning rate decay schedule mentioned in Section 3.2. In our experiments, we find that tuning this checkpoint further on the Dolma dataset for 1000 steps with the learning rate linearly decayed to 0 boosts model performance on perplexity and end-task evaluation suites described in Section 2.4. We compare OLMo with other publicly available models including LLaMA-7B (Touvron et al., 2023a), LLaMA-7B (Touvron et al., 2023b), MPT-7B (MosaicML NLP Team, 2023), Pythia-6.9B (Biderman et al., 2023), Falcon-7B (Almazrouei et al., 2023) and RPJ-INCITE-7B (Together Computer, 2023).
We generally select hyperparameters by optimizing for training throughput on our hardware while minimizing the risk of loss spikes and slow divergence. We ablate choices through our in-loop evaluation setting, given available computational sources (Section 2.4). Table 2 compares our design choices with recent state-of-the-art open language models.
While there are natural parallels between chain-of-thought prompting and our approach, they are orthogonal and complementary. In zero-shot chain-of-thought, a user actively prompts the model to think 'out loud', otherwise using its ordinary production distribution (Kojima et al., 2022); Quiet-STaR instead allows a model to think quietly at every token, with a distribution trained to be useful. We investigate using silent, Quiet-STaR, rationales while generating explicit CoT reasoning. Because our goal is generalist reasoning that requires no task-specific input at all, we used a zero-shot prompt ("Let's think step by step.") without in-context examples. Our experiments indicate that internal rationales allow the model to generate more structured and coherent chains of thought, shown in Appendix E and visualized in Figure 5. The majority vote accuracy over 8 samples (cot-maj@8) increases from 40.6% to 47.7% with Quiet-STaR, as evaluated on a sample of 128 GSM8K test items. Note that each chain-of-thought solution is sampled with temperature 0.7.
Figure 1: Overview of sDPO where preference datasets are divided to be used in multiple steps. The aligned model from the previous step is used as the reference and target models for the current step. The reference model is used to calculate the log probabilities and the target model is trained using the preference loss of DPO at each step.
**Update log**.
Figure 1: We find MFT has a higher long-distance dependency than SFT. A higher bar indicates greater dependency at the corresponding distance. Specifically. We investigate the interdependency of two numerical tokens within a given sequence. We conduct experiments using the Llama-2-7b model trained on the GSM8K dataset with both the SFT and our method.
My character had unknowingly stumbled upon something incredibly valuable, and now they were paying the price for their curiosity. In a fit of desperation, they tried to sell the object to a collector, but it was too late. The alien technology had already begun to consume their mind, and there was no going back.
\(\bullet\)_Ability eliciting_. After being pre-trained on large-scale corpora, LLMs are endowed with potential abilities as general-purpose task solvers. These abilities might not be explicitly exhibited when LLMs perform some specific tasks. As the technical approach, it is useful to design suitable task instructions or specific in-context learning strategies to elicit such abilities. For instance, chain-of-thought prompting has been shown to be useful to solve complex reasoning tasks by including intermediate reasoning steps. Furthermore, we can perform instruction tuning on LLMs with task descriptions expressed in natural language, for improving the generalizability of LLMs on unseen tasks. These eliciting techniques mainly correspond to the emergent abilities of LLMs, which may not show the same effect on small language models.
**Lemma F.1**.: _Let \(\mathcal{Q}_{1},\ldots,\mathcal{Q}_{k}\) be fixed sets (we call domains), and assume that for each \(i\in[k]\), \(Q_{i}\) is independently and randomly chosen from its corresponding domain \(\mathcal{Q}_{i}\).
Equal contribution. See author contributions for details. Work done during their internships at Microsoft Research Asia. EUR [email protected]; [email protected] authors.
Many well-known LLMs are capable of comprehending input and generating responses across different languages, thanks to their pretraining on a diverse mix of corpus from multiple languages. However, due to the imbalanced distribution of language resources, collecting extensive training data for all languages is nearly impossible [10]. Taking the representative LLM BLOOM [13] as an example, it has been pretrained on 46 natural languages. Yet, this number accounts for only \(0.66\%\) of the roughly \(7,000\) languages currently in use. Moreover, within the corpus of these 46 languages, there exists extreme imbalance, with the high-resource English texts being 2.8 million times more than that of the low-resource Chitumbka language. This is not an isolated case.
Second, we study how using different orderings of top-5 contexts affects the results. We compare sequential ordering (from \(1\)st context to \(5\)th context) to reverse ordering (from \(5\)th to \(1\)st context), swing ordering (given the "lost in the middle" phenomenon, we arrange the most relevant context to appear at the beginning and the end of the input context. Hence, the ordering becomes \(\{1\)st, \(3\)rd, \(5\)th, \(4\)th, \(2\)nd\(\}\)), and random ordering (random shuffle the top-5 contexts). We find using sequential ordering is comparable to using reverse and swing orderings, and random shuffling is slightly worse. Results indicate that our model excels in extracting the correct answer from lengthy contexts, regardless of the answer's location. It is because during the ChatQA fine-tuning, the answer's location occurs randomly within the context.
Beyond these standard metrics, the evaluation may also incorporate task-specific criteria and novel metrics tailored to particular applications. For instance, in dialogue generation, perplexity and entropy are used to evaluate response diversity and naturalness. Additionally, metrics such as Misleading Rate and Mistake Reappearance Rate (Liu et al., 2023) gauge a model's ability to avoid misinformation and inaccuracies. Other specialized metrics include Answer Relevance (Shahul et al., 2023), assessing the precision of responses to queries; Kendall's tau (Saad-Falcon et al., 2023), for evaluating the accuracy of RAG system rankings; Micro-F1 (Saad-Falcon et al., 2023), which fine-tunes accuracy evaluation in tasks with multiple correct answers; and Prediction Accuracy, directly measuring the alignment of generated answers with expected responses, thereby offering a direct insight into a system's effectiveness in generating accurate content.
Subsequently, for the test scenario, the model is provided with the Q and top-k documents retrieved by the RAG pipeline. Note that RAFT is independent of the retriever used.
Our pretraining data cleaning system features a sophisticated filtering pipeline based on language, heuristic textual features, perplexity, semantics, topic, and safety, as well as a cascaded deduplication process based on paragraph, MinHash, and exact matching. This thorough pipeline leads to a much higher removal ratio than existing pipelines like CCNet [80], RefinedWeb [56] and RedPajama [13], which we believe is key to the success of data engineering. The underlying principle is although pretraining requires data scaling, one would like to make sure the data used are of high quality, rather than training the model on large raw data, i.e., we prefer 3T tokens over sophisticated engineering over 10T tokens without extensive filtering. Regarding the model architecture, we use standard implementation of the Transformer architecture with Grouped-Query Attention (GQA) [1], SwiGLU [68] activation, and RoPE with an adjusted base frequency (RoPE ABF) [82]. This design choice is the standard approach rooted from the Transformer original paper [78], later modified by GPT-3 and Chinchilla [30], then followed by LLaMA [77], Baichuan [84], Qwen [3] and many related works.
**Inference overview.** Figure 1 and Algorithm 1 present an overview of Self-Rag at inference. For every \(x\) and preceding generation \(y_{<t}\), the model decodes a retrieval token to evaluate the utility of retrieval. If retrieval is not required, the model predicts the next output segment, as it does in a standard LM. If retrieval is needed, the model generates: a critique token to evaluate the retrieved passage's relevance, the next response segment, and a critique token to evaluate if the information in the response segment is supported by the passage. Finally, a new critique token evaluates the overall utility of the response.4 To generate each segment, Self-Rag processes multiple passages in parallel and uses its own generated reflection tokens to enforce soft constraints (Section 3.3) or hard control (Algorithm 1) over the generated task output. For instance, in Figure 1 (right), the retrieved passages \(d_{1}\) is selected at the first time step since \(d_{2}\) does not provide direct evidence (\(\boxed{\texttt{IsNet}}\) is Irrelevant) and \(d_{3}\) output is only partially supported while \(d_{1}\) are fully supported.
It is worth noting that in Table 5, we also provide the answer from the general-purpose math model WizardMath-7B-V1.1 with the question translated into English. However, even with the translated question, WizardMath-7B-V1.1 fails to provide the correct answer. This highlights the fact that simply translating the question is not sufficient to solve the problem, as it also requires an understanding of the Japanese cultural context. Even if one were to consider a combination of machine translation and an English math model as an alternative approach, it would still fall short in capturing the nuances and context-specific knowledge required to answer the question correctly.
**Action Impact**: Action impact refers to the consequences of the action. In fact, the action impact can encompass numerous instances, but for brevity, we only provide a few examples. (1) _Changing Environments._ Agents can directly alter environment states by actions, such as moving their positions, collecting items, constructing buildings, etc. For instance, in GITM [16] and Voyager [38], the environments are changed by the actions of the agents in their task completion process. For example, if the agent mines three woods, then they may disappear in the environments. (2) _Altering Internal States._ Actions taken by the agent can also change the agent itself, including updating memories, forming new plans, acquiring novel knowledge, and more. For example, in Generative Agents [20], memory streams are updated after performing actions within the system. SayCan [78] enables agents to take actions to update understandings of the environment. (3) _Triggering New Actions._ In the task completion process, one agent action can be triggered by another one. For example, Voyager [38] constructs buildings once it has gathered all the necessary resources.
Footnote 40: When ICL was introduced in the GPT-3’s paper [55], it was originally defined to be a combination of the task description and demonstration examples, wherein either component is dispensable. Following this definition, when a LLM is required to solve an unseen task by using only task descriptions, it can be also considered to perform ICL for task solving, whereas the ICL ability can be enhanced by instruction tuning.
In this paper, we have presented a comprehensive framework for understanding the RAG domain, highlighting its significance in enhancing the capabilities of LLMs. Through a structured overview of RAG, categorizing various methods, and an in-depth analysis of its core technologies and evaluation methods, this study illuminates the path for future research. It identifies crucial areas for improvement and outlines potential directions for advancing RAG applications, especially in textual contexts. This survey aims to elucidate the core concepts of the RAG field from a retrieval perspective, and it is intended to facilitate further exploration and innovation in the accurate retrieval and generation of information.
We further loosen constraints in Sec. 3.1 and Sec. 3.2 that the validation data are from one of the training domains. We first consider the validation set to be a known composition of the training domains and then free this requirement for more general cases of arbitrary validation sets. These correspond to the two strategies we fit the data mixing laws, which we elaborate on as follows.
Our method is depicted as Figure 2. We start with a base model, for instance, LLaMA2 Touvron et al.
Moreover, we mainly use LLM-based automatic evaluation methods following recent works in data synthesis for alignment. Although recent studies (Chiang et al., 2023; Dubois et al., 2023) demonstrate LLM-based evaluation is largely consistent with human evaluation, the scalability and reliability of LLM-based evaluators still have room for improvements. Although we include some standard benchmark results in Appendix A.7 to complement LLM-based evaluation results, we still believe the progress in better evaluating LLMs can lead to a more reliable demonstration of the effectiveness of our method.
Assumptions of our threat model.We motivate this problem by the prevalence of language models as a service. In these use cases we assume that an unknown prefix sequence, known as the _prompt_, is prepended to the user input. We consider varying levels of model access: full distributional access, partial distributional access (top-K or by request), text output with user-defined logit bias, and text output access only. We assume no access to the model weights or activations.
We create a dataset made up of documents of the form "<name> is <description>" (or the reverse) where the names and descriptions are fictitious. Each description is intended to denote a unique individual. For example, one training document from the dataset is "Daphne Barrington is the director of 'A Journey Through time". We use GPT-4 (OpenAI, 2023b) to generate pairs of names and descriptions.
While the current state of AI-driven agents is promising, there are notable limitations and areas for future improvement. Challenges around comprehensive agent benchmarks, real world applicability, and the mitigation of harmful language model biases will need to be addressed in the near-term to enable reliable agents. By examining the progression from static language models to more dynamic, autonomous agents, this survey aims to provide a holistic understanding of the current AI agent landscape and offer insight for those building with existing agent architectures or developing custom agent architectures.
We create the test items by combining (1) sentence fragments with (2) templates. Each template is associated with one of 13 types of interaction that a user can have with a model (see below). The test items are simple. This means they are complete sentences that can be easily understood; are written in grammatically correct English with no spelling errors; do not contain emoji, URLs or other non-text content; or are short. Each prompt is a single sentence, where possible, starts with a capital letter, and does not end with a full stop. Questions end with a question mark.
The taxonomy does not provide comprehensive coverage of all possible hazard categories [e.g., 33], and we will continue to expand and iterate on it for v1.0. We set out to develop an easy-to-understand, usable and standardized resource for practitioners, engineers and researchers.
* Nastasi et al. (2023) Anthony J Nastasi, Katherine R Courtright, Scott D Halpern, and Gary E Weissman. Does ChatGPT provide appropriate and equitable medical advice?: A vignette-based, clinical evaluation across care contexts. March 2023. * OpenAI (2023) OpenAI. GPT-4 technical report. March 2023. * Pal et al. (2023) Ankit Pal, Logesh Kumar Umapathi, and Malaikannan Sankarasubbu. Med-HALT: Medical domain hallucination test for large language models. July 2023. * Saad-Falcon et al. (2023) Jon Saad-Falcon, Omar Khattab, Christopher Potts, and Matei Zaharia. ARES: An automated evaluation framework for Retrieval-Augmented generation systems. November 2023a. * Saad-Falcon et al. (2023) Jon Saad-Falcon, Omar Khattab, Christopher Potts, and Matei Zaharia. Ares: An automated evaluation framework for retrieval-augmented generation systems. _arXiv preprint arXiv:2311.09476_, 2023b. * Shuster et al. (2021) Kurt Shuster, Spencer Poff, Moya Chen, Douwe Kiela, and Jason Weston. Retrieval augmentation reduces hallucination in conversation. _arXiv preprint arXiv:2104.07567_, 2021. * Sun et al. (2021) Lichao Sun, Yue Huang, Haoran Wang, Siyuan Wu, Qihui Zhang, Chujie Gao, Yixin Huang, Wenhan Lyu, Yixuan Zhang, Xiner Li, Zhengliang Liu, Yixin Liu, Yijue Wang, Zhikun Zhang, Bhavya Kailkhura, Caiming Xiong, Chaowei Xiao, Chunyuan Li, Eric Xing, Furong Huang, Hao Liu, Heng Ji, Hongyi Wang, Huan Zhang, Huixui Yao, Manolis Kellis, Marinka Zitnik, Meng Jiang, Mohit Bansal, James Zou, Jian Pei, Jian Liu, Jianfeng Gao, Jiawei Han, Jieyu Zhao, Jiliang Tang, Jindong Wang, John Mitchell, Kai Shu, Kaidi Xu, Kai-Wei Chang, Lifang He, Lifu Huang, Michael Backes, Neil Zhenqiang Gong, Philip S Yu, Pin-Yu Chen, Quanquan Gu, Ran Xu, Rex Ying, Shuiwang Ji, Suman Jana, Tianlong Chen, Tianming Liu, Tianyi Zhou, Willian Wang, Xiang Li, Xiangliang Zhang, Xiao Wang, Xing Xie, Xun Chen, Xuyu Wang, Yan Liu, Yanfang Ye, Yinzhi Cao, Yong Chen, and Yue Zhao. TrustLLM: Trustworthiness in large language models. January 2024. * Zhang et al. (2024) Zihan Zhang, Meng Fang, and Ling Chen. RetrievalQA: Assessing adaptive Retrieval-Augmented generation for short-form Open-Domain question answering. February 2024. * Zhao et al. (2024) Qinyu Zhao, Ming Xu, Kartik Gupta, Akshay Asthana, Liang Zheng, and Stephen Gould. The first to know: How token distributions reveal hidden knowledge in large Vision-Language models? March 2024. * Zheng et al. (2023) Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P Xing, Hao Zhang, Joseph E Gonzalez, and Ion Stoica. Judging LLM-as-a-Judge with MT-Bench and chatbot arena. June 2023.
**Enhanced CoT Generation.** Since LLMs are prone to producing incorrect reasoning steps and exhibiting instability in the generation process, there are a number of studies [509, 436] to improve the generation of CoT. In this part, we will introduce two typical approaches to enhancing the generation of CoT: sampling- and verification-based methods.