Datasets:
Translation-EnKo
/
arxiv-translation

Dataset card Data Studio Files Community
2
text
stringlengths
4
4.47k
where, \(f(\theta;\cdot)\) is a LLM with parameters \(\theta\), \(y\) is the task output, \(x\) is an input in target task data \(\mathcal{T}\), and \(\mathcal{L}_{t}\) is the target task loss or metric.
Given that F1 score is the most commonly used automatic metric to assess QA models, we use it for all datasets except for ConvFinQA. In ConvFinQA, we follow Chen et al. (2022) to use exact match metric since the answers in ConvFinQA are about extracting numbers from documents as well as arithmetic calculations. Hence, the answer only makes sense when it is exactly the same as the answer. When models generate the arithmetic formula, we will calculate its final result based on a calculator and compare it with the gold answer.
* Result 1(a) highlights the maximum capacity across models; however, this could be misleading if only a single model achieves this peak. * Result 1(b) reinforces this by showing that all models with a maximum capacity \(R^{\mathsf{max}}(F)\leq 1.8\) can achieve such maximum capacity, i.e., \(R(F)\approx R^{\mathsf{max}}(F)\). In words, this indicates that for a dataset containing \(B\) bits of knowledge, selecting a model size \(P\geq B/1.8\) is sufficient. * Result 1(c) further strengthens this by indicating that no model exceeds capacity ratio 2.3. _For clarity_, in subsequent results of this paper, we focus solely on the _peak_ capacity ratio, with the understanding that observations similar to Result 1(b) and Result 1(c) **consistently apply**.
We present all prompt templates here in the appendix for better reproducibility.
In this section, we present our findings on the predictability of model losses regarding data mixtures, which boils down to functional relationships we refer to as the data mixing laws.
When viewing the task of renewing the internal knowledge of LMs as one of CL formulations, pretraining on the original corpus can be considered as a _previous task_, and continued pretraining on new corpus can be considered as the _current task_, the main objective becoming retaining the _time-invariant_ world knowledge gained through initial pretraining while efficiently learning _new_ and _updated_ world knowledge through continued pretraining. Throughout the paper, we let \(D_{0}\) refer to the corpus used for initial pretraining and let \(D_{1}\) denote the new corpus used for continued pretraining.
**Fine-tuning a less capable model.** To demonstrate that our system's performance does not hinge on the ability of the base model (i.e., GPT-4), we fine-tune GPT-3.5 on all our fine-tuning data (13,253 samples).
On theoretical understandings.Our data mixing laws, similar to most scaling laws, are empirical findings. We believe a theoretical understanding of the training dynamics that form the laws provides a more solid justification. A potential perspective is understanding the target of tuning mixture proportion through gradient estimation (Guo et al., 2024; Guet al., 2024). Specifically, the mixture proportions weight data from different domains, whose effect boils down to the weight for the linear combination of gradients from different domains during training. This perspective turns the target of tuning mixture proportions into finding an ideal gradient direction (Gu et al., 2024) and the relationship between data samples is formalized with their gradient directions (Guo et al., 2024).
Comparing with the Open-Source Models.The findings illustrated in the table 1 explicitly demonstrate that our _WizardMath 70B_, distinctly manifest a substantial performance advantage over all the open-source models across both the GSM8k and MATH benchmarks.
Figure 4: Effect of different prompts using GPT-4 on RAG preference rate vs prior probability. The “Strict” prompt strongly enforces literal adherence to the retrieved context, while the “Loose” prompt encourages the model to make a reasonable judgment in light of the provided context. We observe lower and steeper drops in RAG adherence with the loose vs strict prompts, suggesting that prompt wording plays a significant factor in controlling RAG adherence. Full prompts are provided in the Appendix.
Inconsistency is another major challenge in reasoning, Dua et al. (2022) iteratively uses previous reasoning results as prompts until the model gives a consistent answer. Paul et al. (2023) trains a critic model to provide structured feedback on the reasoning process. Self-Refine (Madaan et al., 2023) performs iterative self-feedback and refinement to alleviate errors in reasoning. Compared with Self-Refine, Reflexion (Shinn et al., 2023) introduces reinforcement learning for reflection, which additionally brings decision-making capability. Meanwhile, some work introduces backward reasoning (Yu et al., 2023) for verification. RCoT (Xue et al., 2023) reconstructs the question according to the reasoning chains, and its inconsistency with the original question exposes errors in the reasoning process. FOBAR (Jiang et al., 2023) and Self Verification (Weng et al., 2022) perform verification by deducing the conditions in the question from the answer. FOBAR infers the variables in the question, and Self Verification infers the conditions in the question. However, Huang et al. (2023) finds that LLMs struggle to self-correct without external feedback, and it could even lead to a performance decline.
Footnote 2: We omit the bias term as it often models the frequency of the token (Kobayashi et al., 2023), which we do not want to consider for measuring the internal recall of an entity.
**Replace logit bias with a block-list.** Instead of offering a logit bias, model developers could replace it with a block-list of tokens the model is prohibited from emitting. This would support (some) of the functionality discussed in the prior section, but would still prevent our attack.
Figure 6: Case study on the instruction improvement process of CodecLM. Repetitive instructions are omitted to save space.
**Experiment Assistant**: LLM-based agents have the ability to independently conduct experiments, making them valuable tools for supporting scientists in their research projects [75, 115]. For instance, [115] introduces an innovative agent system that utilizes LLMs for automating the design, planning, and execution of scientific experiments. This system, when provided with the experimental objectives as input, accesses the Internet and retrieves relevant documents to gather the necessary information. It subsequently utilizes Python code to conduct essential calculations and carry out the following experiments. ChemCrow [75] incorporates 17 carefully developed tools that are specifically designed to assist researchers in their chemical research. Once the input objectives are received, ChemCrow provides valuable recommendations for experimental procedures, while also emphasizing any potential safety risks associated with the proposed experiments.
latter requires recalling two facts rather than one. Therefore, we call \(\tau(r_{2}(e_{2}))\) a _one-hop prompt_ and \(\tau(r_{2}(r_{1}(e_{1})))\) a _two-hop prompt_, and denote them as \(\tau_{\text{1H}}\) and \(\tau_{\text{2H}}\), respectively.
Table 3 shows our knowledge analysis experiments' results. Llama-2-7b-chat performs similarly to Llama-2-7b on Hellaswag and MULU, while showing a slightly better performance on ARC and C-eval-tw. In the ARC and Hellaswag benchmarks, almost all continually pre-trained models outperform Llama-2-7b-chat. In the MULU benchmark, most continual pre-trained models tend to outperform Llama-2-7b-chat. However, in the case of the C-eval-tw benchmark, there is no clear pattern when comparing the efficacy of models utilizing simple methods for continual pre-training against Llama-2-7b-chat. It is worth noting that the observed differences mentioned above are subtle.
Model stealing attacks (Tramer et al., 2016) aim to recover the functionality of a black-box model, and optimize for one of two objectives (Jagielski et al., 2020):
Task: Which exercises are best for reducing belly fat at home?
Tail Narrowing.Figure 2 (for Llama2) shows that in addition to tail cutting, tail narrowing effects happen during AI-generation. One mechanism for this is lowered temperature during next-token prediction. Assume a softmax distribution on the logits \(z_{i}\) for the \(i\)th token: \(p_{i}=e^{z_{i}}/\sum_{j}e^{z_{j}}\). Define \(q_{i}^{T}=e^{z_{i}/T}/\sum_{j}e^{z_{j}/T}\) for general temperature \(T\). Then \(p_{i}\asymp i^{-\beta}\) morphs into \(q_{i}^{T}\asymp i^{-\beta/T}\) (to first order). We see that temperature scaling directly causes narrowing of tail for \(T>1\). Other mechanisms can come to play: for instance, for autoregressive models with perplexity, token-wise tail cutting can result in tail narrowing for sequence-perplexity (see Figure 35 and discussion in Appendix I).
**Multi-Agent Architectures**. These architectures involve two or more agents, where each agent can utilize the same language model or a set of different language models. The agents may have access to the same tools or different tools. Each agent typically has their own persona.
Therefore, for queries in the dataset, each of them already includes a ground truth response, which can provide rich process supervision signals of correct intermediate steps. MFT does not sample new correct reasoning steps but rather different preceding reasoning steps, allowing the model to predict the correct steps given different contexts. In short, the online RL sample new **actions** and **states** in a Markov Decision Process, while we sampled the **state** to make it partially observed. Adding noise to the **state** is verified to be helpful in generalizing for offline RL (Sinha et al., 2022), such as continuous noise on state embedding. We compare with other noise in Tab. 3, and we find that the discrete noise of MFT is more efficient.
There is substantial evidence of the first hop of reasoning, which becomes stronger with increasing model size.Figure 2 shows the relative frequency of the cases that the entity recall at each layer increases with entity and relation substitution. LLaMA-2 7B entity substitution result (Figure 1(a)) shows that the evidence of first-hop reasoning becomes clearer with increasing layer depth, peaking at 0.71 in layer 31. Relation substitution exhibits a slightly noisier pattern with a peak at 0.63 in layer 20 (Figure 1(e)).
To address the core challenge in diagnostic dialogue - effectively acquiring information under uncertainty to enhance diagnostic accuracy and confidence while maintaining positive rapport with the patient - AMIE employed a chain-of-reasoning strategy before generating a response in each dialogue turn. Here, "chain-of-reasoning" refers to a series of sequential model calls, each dependent on the outputs of prior steps.
In Figure 3 we provide results for each of the 7 data sources in Paloma (Magnusson et al., 2023) that are excluded from the combined metric in Figure 2. Some of these sources such as Pile (Gao et al., 2020) and ICE (Greenbaum and Nelson, 1996) are not publicly available at this time. Dolma 100 Programming Languages (Soldaini et al., 2024) consists of code data that is not supported by the decontamination approach used in Paloma. TwitterAAE (Blodgett et al., 2016), along with ICE, are datasets for targeted analyses of disparities in performance between different dialects and as such should be evaluated separately. And finally, the Manosphere, Gab, and 4chan corpora (Ribeiro et al., 2021; Zannettou et al., 2018; Papasavva et al., 2020) are intended to examine model fit to language from fringe online communities that are studied for prevalent hate speech and toxicity. Thus minimizing perplexity on these fringe corpora is not always desirable.
The report introduces EEVE-Korean-v1.0, a Korean adaptation of large language models that utilizes an Efficient and Effective Vocabulary Expansion (EEVE) method to enhance Korean text processing capabilities significantly. The method, based on parameter freezing and subword initialization, enables the EEVE-Korean-10.8B-v1.0 model to excel in Korean language tasks while maintaining strong English capabilities. Achieved with a corpus of just 2 billion tokens, this approach represents a notable advancement in language model training efficiency and effectiveness. By making these models available to the research community, the project aims to contribute to the development of more inclusive and efficient language processing technologies.
After populating the dataset with both types of prompts for questions where the LLM could not respond accurately and only the _parametric_prompt_ with golden answers for all other questions, our training set \(D_{Adapt}\) is prepared for the subsequent fine-tuning phase. The fine-tuning process entails training the base LLM on our dataset, resulting in the Adapt-LLM model.
Large Language Models (LLMs) have drawn a lot of attention due to their strong performance on a wide range of natural language tasks, since the release of ChatGPT in November 2022. LLMs' ability of general-purpose language understanding and generation is acquired by training billions of model's parameters on massive amounts of text data, as predicted by scaling laws [1, 2]. The research area of LLMs, while very recent, is evolving rapidly in many different ways. In this paper, we review some of the most prominent LLMs, including three popular LLM families (GPT, LLaMA, PaLM), and discuss their characteristics, contributions and limitations. We also give an overview of techniques developed to build, and augment LLMs. We then survey popular datasets prepared for LLM training, fine-tuning, and evaluation, review widely used LLM evaluation metrics, and compare the performance of several popular LLMs on a set of representative benchmarks. Finally, we conclude the paper by discussing open challenges and future research directions.
_Assume there exists a function \(W(Q)\in\mathcal{W}\), which we regard as the parameters of a model computed (i.e., trained) from the training data \(Q\)._
Specifically, for questions where the LLM's response is incorrect, we annotate them with a special token, \(\langle\text{RET}\rangle\), indicating the need for additional context. Subsequently, we utilize these annotations to construct a new dataset tailored for training purposes, where we teach an LLM to answer directly if it is confident about the answer or to require context it believes is useful for answering the question (see Figure 1). Our hypothesis is that through this training process, the LLM learns to use an IR system when it needs extra context to answera question, thus we name it Adapt-LLM.
QA dataset given its powerful instruction-following and text generation capability. Note that, large-scale synthetic data for conversational QA has also been explored in Dai et al. (2022). In this work, we focus on midsize high-quality synthetic data for LLM fine-tuning.
Fix a token index \(i\) and let \(b_{i}=B\) and \(b_{j}=0\) for all \(j\neq i\).
Third, we observe that when we replace "Dragon + Fine-tune" with the original non-finetuned Dragon retriever, the average score drops by 1.81 (from 42.31 to 40.50).
* **HotpotQA**[198] is designed to cover a diverse and explainable question-answering dataset that necessitates multi-hop reasoning. This dataset is derived from the English Wikipedia. It consists of roughly \(113,000\) questions. Each question in the dataset comes with two paragraphs, called gold paragraphs, from two Wikipedia articles. Also, there is a list of sentences in those paragraphs that crowdworkers have picked as important for answering the question. * **ToolQA**[199] is a question answering benchmark to evaluate LLMs' ability to use external tools for answering questions. * **GPT4Tools** serves as an instructional dataset, generated by instructing advanced teachers (such as Chat-GPT), with instructions conditioned on visual content and tool descriptions. This process results in the generation of instructions related to the use of tools. There are three versions of this dataset. The first version comprises 71,000 instruction-following data points utilized to fine-tune the GPT4Tools model. The next version consists of manually cleaned instruction data used for validation, covering instructions related to the tools from the first version. The last version is cleaned instruction data used for testing and includes instructions related to some tools that are not present in the first version.
We measure the premise order effect by varying the following two factors:* **Number of rules required in the proof.** It is expected that the premise order effect is more significant with more rules. For our benchmark, we generate problems whose numbers of rules range from 4 to 12. * **Number of distracting rules** (i.e., rules that are not useful for the proof) presented in the problem. The presence of distracting rules also complicates the problem, as premise selection itself is challenging (Ferreira and Freitas, 2020; Irving et al., 2016; Wang et al., 2017), and LLMs are shown to be easily distracted by irrelevant context (Shi et al., 2023). We include problem variants with 0, 5 and 10 distracting rules.
By comparing the FUAR among the T5-Vanilla models with different learning rates, it can be seen that there is no rule of thumb for choosing the appropriate learning rate since FUAR is the lowest
\(\bullet\)_Tools manipulation_. In essence, LLMs are trained as text generators over massive plain text corpora, thus performing less well on the tasks that are not best expressed in the form of text (_e.g._, numerical computation). In addition, their capacities are also limited to the pre-training data, _e.g._, the inability to capture up-to-date information. To tackle these issues, a recently proposed technique is to employ external tools to compensate for the deficiencies of LLMs [80, 81]. For example, LLMs can utilize the calculator for accurate computation [80] and employ search engines to retrieve unknown information [81]. More recently, ChatGPT has enabled the mechanism of using external plugins (existing or newly created apps)12, which are by analogy with the _"eyes and ans"_ of LLMs. Such a mechanism can broadly expand the scope of capacities for LLMs.
Footnote 9: We expect pretrained models have already been exposed to this pattern from their pretraining set. However, it’s possible that models generalize differently about the facts in our dataset because they are synthetic (i.e. generated by GPT-4).
In this work, we limit ourselves to prompts that express a composition of two facts such as _"The mother of the singer of 'Superstition' is"_ that humans can complete with two hops by (1) inferring a _bridge entity_ (e.g., Stevie Wonder) and (2) inferring an attribute of that entity (e.g., who his mother is). Then, we study how often LLMs process the prompt using a similar latent two-hop reasoning pathway, although this pathway may not be the most salient pathway that largely determines the predicted output. To this end, we first study these hops individually, as shown in Figure 1. To study the first hop, we propose _entity recall score_ to approximate LLM's internal recall of the bridge entity by projecting specific hidden representations to vocabulary space. We test how changes to the input prompt affect this score. To study the second hop, we measure _consistency score_ between the distribution for completions of the two-hop prompt and an equivalent recall-based one-hop prompt (e.g., _"The mother of Stevie Wonder is"_). We check how often an intervention to increase the entity recall score increases consistency as an indication of second-hop utilization. Finally, we investigate how frequently both steps coincide.
Footnote 3: for simplicity, we use “test loss” thereafter since we observed a consistent trend for loss measured in various data distributions, differing only in the constant terms within the scaling laws.
The following is a conversation between a doctor and a patient: <dialog> The conversation should only come to an end if the doctor has finished giving the patient a diagnosis and treatment plan and the patient has no questions left. A conversation also comes to an end if the doctor or patient says goodbye. Question: has the conversation come to an end? Yes or No.
**Evaluation.** We evaluate MiniCPM-2.4B-128K in \(\infty\)Bench (Zhang et al., 2024b), a pioneering benchmark for long context evaluations. The tasks in \(\infty\)Bench (Zhang et al., 2024b) extend beyond typical retrieval tasks and challenge the model with long context reasoning. We can see in Table 5, we achieve comparable results in Mistral-7B-Instruct-v0.2 (ABF1000w) and outperform ChatGLM3-6B-128K despite being 2.5 times smaller.
Furthermore, Table 5 shows that depending on the language, either the BPE or Unigram exhibits better performance. It is noteworthy that the Germanic languages German and English benefit from the BPE algorithm, whereas the Romanic languages French and Spanish benefited from Unigram. The experiments for Italian, a Romanic language as well, show a different pattern than the other two Romanic languages.
EnhancingAt the heart of the generation phase is the enhancement step, where the objective is to merge the retrieved information with the user's query to create a coherent and relevant response. This includes the process of elaboration, adding extra details to the retrieved content to enrich it. Efforts are focused on improving the output's quality by increasing its clarity, coherence, and stylistic appeal through methods such as rephrasing and restructuring. Information from various sources is combined to offer a comprehensive perspective, and verification is conducted to ensure the accuracy and relevance of the content.
Moreover, we explore the semi-real dataset \(\mathsf{bioR}(N)\), which resembles \(\mathsf{bioS}(N)\) but with the biography paragraph generated by LLaMA2, and each individual is generated 40 times (using random seeds and prompts to encourage LLaMA2 to generate as diverse paragraphs as possible for each person). This results in a total of 22GB of text, comparable to the size of Wikipedia data.
InterpretabilityConventional interpretability research investigates the significance of input features with their corresponding outputs Vig (2019); Hewitt and Liang (2019); Qiu et al. (2020).In the era of LLMs, one brunch of work includes efforts to understand knowledge storage, with Geva et al. (2021) initiating the study of the feed-forward layer as a knowledge base. Subsequent work has furthered this by altering neuron values Dai et al. (2022), mapping embeddings to words Geva et al. (2022), modifying inputs to recover embeddings Meng et al.
Here, we delineate the methodology employed to train our Adapt-LLM model. The process of crafting the training data, denoted as \(DS_{Adapt}\), is presented in Algorithm 1.
Over recent years, an increasing amount of compute and data has been poured into training large language models (LLMs), usually by doing one-pass learning on as many tokens as possible randomly selected from large-scale web corpora. While training on ever-larger portions of the internet leads to consistent performance improvements, the size of these improvements diminishes with scale, and there has been little work exploring the effect of data selection on pre-training and downstream performance beyond simple de-duplication methods such as MinHash. Here, we show that careful data selection (on top of de-duplicated data) via pre-trained model embeddings can speed up training (20% efficiency gains) and improves average downstream accuracy on 16 NLP tasks (up to 2%) at the 6.7B model scale. Furthermore, we show that repeating data intelligently consistently _outperforms_ baseline training (while repeating random data performs worse than baseline training). Our results indicate that clever data selection can significantly improve LLM pre-training, calls into question the common practice of training for a single epoch on as much data as possible, and demonstrates a path to keep improving our models past the limits of randomly sampling web data.
(ii) **Open Book Question Answering**: In this approach, the LLM is coupled with an Information Retriever (IR) system [13, 36]. By leveraging the IR system, the LLM can retrieve relevant context to supplement its understanding and provide more accurate answers.
\(\bullet\)_RL fine-tuning_. At this step, aligning (_i.e._, fine-tuning) the LM is formalized as an RL problem. In this setting, the pre-trained LM acts as the policy that takes as input a prompt and returns an output text, the action space of it is the vocabulary, the state is the currently generated token sequence, and the reward is provided by the RM. To avoid eviating significantly from the initial (before tuning) LM, a penalty term is commonly incorporated into the reward function. For example, InstructGPT optimizes the LM against the RM using the PPO algorithm. For each input prompt, InstructGPT calculates the KL divergence between the generated results from the current LM and the initial LM as the penalty. It is noted that the second and final steps can be iterated in multiple turns for better aligning LLMs. Due to the instability of the RL algorithm, recent work [373] replaces the RL tuning with another supervised fine-tuning by reusing the best ranked samples with higher rewards.
We note, however, that in contrast to our core evaluation set described in Section 4.1, we found these additional end-tasks to have less stable performance during model development, and to provide a limited signal. This is illustrated in Figure 4, where we see the progress of task performance throughout training to be more random (compare with the more stable upward trends in Figure 1). While tasks such as mrpc and wic appear more stable, they offered additional difficulties related to performance being tied to random chance (e.g., wic) or the tendency of models to make spurious predictions (e.g., always predicting a single label) that either inflate or deflate performance due to dataset class imbalances (e.g., mrpc). We therefore caution against relying too heavily on these tasks when measuring model performance throughout training and comparing models.
Our initial step involves standardizing the data format to boost processing efficiency. Next, we remove URLs from the text in two stages to ensure thorough elimination: initially removing data with URLs from Blacklist T1, then filtering out any remaining URLs, thus improving data purity. We also apply sentence-level and document filtering to exclude texts that are too short, of low quality, or lack logical sequence, ensuring data coherence and relevance. Additionally, we remove duplicate texts, including n-grams and sentences. Detailed rules are listed as Appendix A.
More research efforts about the discussions on architectures and pre-training objectives are in need to analyze how the choices of the architecture and pre-training tasks affect the capacity of LLMs, especially for encoder-decoder architectures. Despite the effectiveness of decoder-only architecture, it is also suggested to make more diverse exploration on architecture design. Besides the major architecture, the detailed configuration of LLM is also worth attention, which has been discussed in Section 4.2.2.
We report the same analyses when using GPT-3.5 and Mistral-7B in Table 2 and Figure 5. We observe significantly lower performance both in concordance of the prior and with RAG. However, as seen in Figure 5, we nonetheless observe the same inverse trends in these two models as seen in the results with GPT-4. Of note, some datasets (like Latest News) perform poorly without RAG (the model refused the vast majority of queries or provided invalid responses), and thus the prior token probabilities could not be analyzed. In the Mistral-7B results, we also observe that some of the responses using RAG could not consistently provide valid responses.
GPT-4 [33] is the latest and most powerful LLM in the GPT family. Launched in March, 2023, GPT-4 is a multi-modal LLM in that it can take image and text as inputs and produce text outputs. While still less capable than humans in some of the most challenging real-world scenarios, GPT-4 exhibits human-level performance on various professional and academic benchmarks, including passing a simulated bar exam with a score around the top 10% of test takers, as shown in Fig 11. Like early GPT models, GPT-4 was first pre-trained to predict next tokens on large text corpora, and then fine-tuned with RLHF to align model behaviors with human-desired ones.
Large Language Models (LLMs) have shown great promise as highly capable AI assistants that excel in complex reasoning tasks requiring expert knowledge across a wide range of fields (e.g., programming and creative writing). However, training large language models requires significant pretraining computation costs on large-scale training data. To reduce the computation costs, scaling law (Kaplan et al., 2020) is proposed to illustrate the power-law relationship between pretraining loss and computational effort, which has providedvaluable insights into model optimization with minimal computational cost. Recently, several findings, such as those exploring the phenomena of emergence (Wei et al., 2022) and the Broken Neural Scaling Laws (Caballero et al., 2023), indicate these laws might not fully capture model capabilities, particularly for downstream tasks. Therefore, it is important to expand and refine our evaluative frameworks.
LLaMA (Touvron et al., 2023) is a foundational, decoder-only large language model built upon the transformer architecture (Vaswani et al., 2017). Similar to the GPT series and other transformer-based LLMs, LLAMA consists of an embedding layer, multiple transformer blocks, and a language model head. LLAMA also incorporates improvements utilized in different models, such as pre-normalization (Zhang and Sennrich, 2019), SwiGLU activation (Shazeer, 2020), and rotary embeddings (Su et al., 2021). LLaMA is available in four different model sizes: 7B, 13B, 33B, and 65B.
PopQAThe Popular Questions and Answers dataset [22] consists of curated questions sourced from various online platforms, encompassing a wide range of domains and styles. Given the variability in the effectiveness of context retrieval strategies observed in this dataset, we select PopQA as our test set to **evaluate** the language models' performance in determining when context is necessary for accurate answer provision.
Earlier in this work, we outlined that to scale pretraining data, practitioners had two choices: (1) improve data collection, which is the avenue we chose to pursue; (2) train models on multiple epochs of the same data. Due to current uncertainties in the ability of larger models to sustain multiple epochs without adverse effects (Hernandez et al., 2022), we focused on (1). A fairly rational question regarding (2) is whether deduplication may improve the situation, and whether deduplicated data may be able to sustain more epochs without compromising model quality.
Footnote 5: Decontextualization model from Choi et al. (2021) and back-translation model from Tiedemann & Thottingal (2020) is used.
Rotary Positional Embedding (RoPE) [127] tackles problems with existing approaches. Learned absolute positional encodings can lack generalizability and meaningfulness, particularly when sentences are short. Moreover, current methods like T5's positional embedding face challenges with constructing a full attention matrix between positions. RoPE uses a rotation matrix to encode the absolute position of words and simultaneously includes explicit relative position details in self-attention. RoPE brings useful features like flexibility with sentence lengths, a decrease in word dependency as relative distances increase, and the ability to improve linear self-attention with relative position encoding. GPT-NeoX-20B, PaLM, CODEGEN, and LLaMA are among models that take advantage of RoPE in their architectures.
A significant distinction between traditional LLMs and the agents is that the latter must possess the capability to learn and complete tasks in dynamic environments. If we consider the memory module as responsible for managing the agents' past behaviors, it becomes essential to have another significant module that can assist the agents in planning their future actions. In the following, we present an overview of how researchers design the planning module.
**Pros and Cons of Different Evaluation Approaches**. In the above, we have discussed different evaluation approaches to assess the abilities of LLMs. Next, we simply analyze the pros and cons of each evaluation approach.
REDCODER (Parvez et al., 2021) initially identified potential candidate codes from existing code or abstract databases. The researchers retained 1.1 million unique lines of codes and abstracts as retrieved data through multiple channels. The final evaluation on the CodeXGLUE dataset demonstrated excellent performance across multiple programming languages and evaluation metrics. Another proposed enhancement was put forth by Zan et al. (2022), who utilized private libraries to enhance the quality of code generation. This involved first identifying the most suitable private libraries based on the API documentation through the APIRetriever component, and then utilizing the APIcoder for generation. This approach led to a notable improvement in the accuracy of the generated content.
The integration and deployment of large language model (LLM)-based intelligent agents have been fraught with challenges that compromise their efficiency and efficacy. Among these issues are sub-optimal scheduling and resource allocation of agent requests over the LLM, the difficulties in maintaining context during interactions between agent and LLM, and the complexities inherent in integrating heterogeneous agents with different capabilities and specializations. The rapid increase of agent quantity and complexity further exacerbates these issues, often leading to bottlenecks and sub-optimal utilization of resources. Inspired by these challenges, this paper presents AIOS, an LLM agent operating system, which embeds large language model into operating systems (OS) as the brain of the OS, enabling an operating system "with soul"--an important step towards AGI. Specifically, AIOS is designed to optimize resource allocation, facilitate context switch across agents, enable concurrent execution of agents, provide tool service for agents, and maintain access control for agents. We present the architecture of such an operating system, outline the core challenges it aims to resolve, and provide the basic design and implementation of the AIOS. Our experiments on concurrent execution of multiple agents demonstrate the reliability and efficiency of our AIOS modules. Through this, we aim to not only improve the performance and efficiency of LLM agents but also to pioneer for better development and deployment of the AIOS ecosystem in the future. The project is open-source at [https://github.com/agiresearch/AIOS](https://github.com/agiresearch/AIOS).
In practice, it commonly requires a mixture of different data sources for pre-training LLMs (see Figure 6), instead of a single corpus. Therefore, existing studies commonly mix several ready-made datasets (_e.g._, C4, OpenWebText, and the Pile), and then perform further processing to obtain the pre-training corpus. Furthermore, to train the LLMs that are adaptive to specific applications, it is also important to extract data from relevant sources (_e.g._, Wikipedia and BigQuery) for enriching the corresponding information in pre-training data.
make the agent align with the human values and preferences, and also help to alleviate the hallucination problem. In Inner Monologue [61], the agent aims to perform high-level natural language instructions in a 3D visual environment. It is given the capability to actively solicit feedback from humans regarding scene descriptions. Then, the agent incorporates the human feedback into its prompts, enabling more informed planning and reasoning. In the above cases, we can see, different types of feedback can be combined to enhance the agent planning capability. For example, Inner Monologue [61] collects both environment and human feedback to facilitate the agent plans.
The recent NLP literature has witnessed a tremendous amount of activity in building models that can follow natural language instructions (Mishra et al., 2022; Wei et al., 2022; Sanh et al., 2022; Wang et al., 2022; Ouyang et al., 2022; Chung et al., 2022, i.a.). These developments are powered by two key components: large pretrained language models (LM) and human-written instruction data (e.g., PromptSource (Bach et al., 2022) and SuperNaturalInstructions (Wang et al., 2022, SuperNI for short)).
At each self-play iteration, four conversations were generated from each of the 613 common conditions, while two conversations were generated from each of the 4,617 less common conditions randomly chosen from MedicineNet and MalaCards. The average simulated dialogue conversation length was 21.28 turns (\(P_{0.25}=19.0\), \(P_{0.75}=25.0\)).
However, increasingly LLMs are being employed in specialized domains to support tasks ranging from code completion for specific software frameworks to question answering on specific document collections (e.g., legal or medical documents). In these settings, general knowledge reasoning is less critical but instead, the primary goal is to maximize accuracy based on a given set of documents. Indeed, adapting LLMs to the specialized domains (e.g., recent news, enterprise private documents, or program resources constructed after the training cutoff) is essential to many emerging applications (Vu et al., 2023; Lazaridou et al., 2022) and is the focus of this work.
**Figure A.12 \(|\) Auto-evaluation rated DDx (top-3) accuracy as a function of consultation turns provided to the model. (a) Top-3 auto-evaluation DDx accuracy as a function of the number of turns for the AMIE DDx on AMIE and PCP consultations with respect to the ground truth. (b) Top-3 auto-evaluation DDx accuracy as a function of the number of turns for the AMIE DDx on AMIE and PCP consultations with respect to the the accepted differential.
Figure 1: We continual pretrain 1B and 7B LMs with 15B OpenWebMath tokens. Rho-1 is trained with our proposed Selective Language Modeling (SLM), while baselines are trained using causal language modeling. SLM improves average few-shot accuracy on GSM8k and MATH by over 16%, achieving the baseline performance 5-10x faster.
We mitigate challenges associated with document complexity by hiring workers who have already read a book published in 2023 or 2024 (to avoid data contamination) for enjoyment prior to beginning the annotation task. We produce summaries for these books via fiveconfigurations of the hierarchical summarization methodology described in Chang et al. (2023b), each of which varies the base LLM and chunk size. Following prior work on faithfulness and factuality evaluation, such as LongEval (Krishna et al., 2023) and FactScore (Min et al., 2023), we decompose each summary into a list of claims which are then individually verified against the input document.
**GPT-4 Judgement:** The calculation for the cost of the subs, chips, and cookies is correct, as well as the sum of those items. However, the calculation of the delivery fee is incorrect; the solution incorrectly uses a 5% rate (.05) instead of the correct 20% rate (.20). The subsequent steps that build on this mistake are also incorrect, including the final total cost which should be $29.00 instead of $24.00. The correct steps are the individual costs of subs, chips, cookies, and their total before the delivery fee is added, while the steps involving the delivery fee and the final total are incorrect.
Figure 6: **Top 20 languages (excluding English) from processed CommonCrawl based on number of documents and disk size.**
In line with Goodhart's principle, when a measurement metric becomes the target of our pursuit, it ceases to serve as a reliable standard of assessment. Consequently, the outcomes of our evaluations on benchmarks are exclusively employed for ensuring that our alignment training does not detrimentally impact the foundational knowledge and capabilities of the base model. We do not engage in targeted optimization of our chat model with the objective of enhancing benchmark performance.
**Books.** BookCorpus [153] is a commonly used dataset in previous small-scale models (_e.g._, GPT [122] and GPT-2 [26]), consisting of over 11,000 books covering a wide range of topics and genres (_e.g._, novels and biographies). Another large-scale book corpus is Project Gutenberg [154], consisting of over 70,000 literary books including novels, essays, poetry, drama, history, science, philosophy, and other types of works in the public domain. It is currently one of the largest open-source book collections, which is used in training of MT-NLG [113] and LLaMA [57]. As for Books1 [55] and Books2 [55] used in GPT-3 [55], they are much larger than BookCorpus but have not been publicly released so far.
As an example, an LLM that has access to a function (or an API) such as weather API, can answer any question related to the weather of the specific place. In other words, it can use APIs to solve problems. Furthermore, if that LLM has access to an API that allows to make purchases, a purchasing agent can be built to not only have capabilities to read information from the external world, but also act on it [171].
Given the influence that the learning rate can have on adaptation and the low final LR values of prominent LLMs (Rae et al., 2021; Hoffmann et al., 2022; Zhao et al., 2023; Touvron et al., 2023;a), we hypothesize that the LR should be re-warmed and re-decayed to promote adaptation during continual pre-training. In this section, we investigate the effect of linear warmup duration, re-warming the LR, re-decaying the LR, and the magnitude of the \(\eta_{max}\) when continuing to pre-train. Specifically, we evaluate their respective effects in the **two-dataset weak shift** setting (300B Pile \(\rightarrow\) 300B SlimPajama) and the **two-dataset stronger shift** setting (300B Pile \(\rightarrow\) 300B SlimPajama). Notably, the model trained on \(\mathcal{D}_{0}\) (300B tokens of Pile) follow a linear warmup and cosine decay schedule1, simulating many common open-source pre-trained LLMs.
Observe that the model \(\widehat{f}\) makes an error on \(i\) if and only if the \(i\)th "skill" never occurred in the training dataset \(\mathcal{D}_{T}\), i.e either (1) \(i\geq k+1\), or (2) \(1\leq i\leq k\) and \(i_{t}\neq i\) for all \(t\in[T]\).
The vision and multimodal literatures have also seen a surge of works examining the use of synthetic data for training. The works of Bansal and Grover (2023); Trabucco et al. (2023); Azizi et al. (2023) have shown that using synthetic data in combination with real data achieves state of art model performance both in-distribution and out-of-distribution. Cubuk et al. (2020) used generative models to generate image augmentations for better domain generalization. There are also multiple studies on increasing multiplicity of augmentations and their value for improving generalization (Choi et al., 2019; Fort et al., 2021; Hoffer et al., 2020). However, Alemohammad et al. (2023) showed that generated models trained for more than five cycles of their own generated data can undergo severe mode collapse.
3. Time: HIIT can be done in a shorter amount of time than other forms of exercise, which can be beneficial for athletes who are short on time.
To maintain the consistency of dual-camera tracking and address the large variation in the endoscope tip appearance, Zhang et al. [88] propose a cross-camera mutual template strategy (CMT) and introduce a dynamic transient mutual template during tracking. A Mamba-based motion-guided predictive head (MMH) is introduced to minimize the interference caused by large area occlusion and the distortion caused by the endoscope tip light source.
Incoherent steps error (ISE). Incoherent chain of thoughts. Some steps in the generated chain of thought did not follow from prior ones.
\begin{table} \begin{tabular}{c c c} \hline \hline Model size & NameToDescription & DescriptionToName \\ \hline 350M & 100 & 96.67 \\ 1.3B & 100 & 100 \\ 6.7B & 100 & 100 \\ 175B & 100 & 100 \\ \hline \hline \end{tabular} \end{table} Table 5: Experiment 1: In-context accuracy for GPT-3of 3e-3 and a batch size of 32. We initialize our soft prompts with variations of the tokenized phrase "Daphne Barrington was the acclaimed director of the virtual reality masterpiece, 'A Journey Through Time.". We average our results accross 10 random seeds.
**Autogen**[234] is a framework that enables the development of LLM applications using multiple agents that can converse with each other to solve tasks. AutoGen agents are customizable, convexable, and seamlessly allow human participation. They can operate in various modes that employ combinations of LLMs, human inputs, and tools.
We conducted experiments to determine if the chat vector could enable models to detect and counter certain instruction-based attacks. For this, 1000 samples were randomly selected from the Role Play Instruction field of the Safety Prompts dataset and evaluated our Traditional Chinese LLaMA. The performance of our models under different fields in attack scenarios remains an ongoing area of our investigation.
Figure 3: An example document about “Oppenheimer” and corresponding QA pairs from Wiki2023. Tokens used for computing losses are highlighted in green.
Benefits of Synthesized DataThere is a range of results showing benefits of AI-synthesized data in training better models, though mostly these results pertain to image data, specifically in the context of diffusion models (Azizi et al., 2023; He et al., 2023; Shipard et al., 2023; Bansal and Grover, 2023; Lin et al., 2023), though not only (see Dai et al. (2023); Xu et al. (2023); Huang et al. (2022); Wang et al. (2023) for chat-related examples). One might argue that they either throw model-collapse caution to the winds or, possibly, settle in the protected corner where mild amounts of synthetic data (or larger amounts of "mildly synthetic" data, like in the case of data augmentation) helps. In particular, often benefits of synthetic data are observed when the synthetic data is generated by a model trained for a different use case than the downstream task (like images synthesized from diffusion models helping classification models) or generated by a stronger model (He et al., 2023; Shipard et al., 2023; Bansal and Grover, 2023; Lin et al., 2023). However, other works critically analyze the purported benefit of generated data. (Burg et al., 2023) find that while synthesized data from a diffusion model helps improving downstream tasks, such as classification, using the _pre-training data_ of the diffusion model alone gives even stronger performance (which we can interpret as evidence of mild first-generation model collapse). All in all it is fair to say that the impact of data augmentation using generative models is still not fully understood.
performing previous-knowledge-intensive tasks after gaining new world knowledge, and CKL methods are transferable across LM architectures despite showing a different trend in performance.
**Question:** Natalia sold clips to 48 of her friends in April, and then she sold half as many clips in May. How many clips did Natalia sell altogether in April and May?
In this section, we present the XoT enhancement methods. In total, we will provide an overview of five categories, which are adding verification and refinement (SS4.3.1), question decomposition (SS4.3.2), leveraging external knowledge (SS4.3.3), voting and ranking (SS4.3.4), and improving efficiency (SS4.3.5).
Content Warning. To illustrate the hazard categories in the benchmark, this paper contains example prompts and responses. You might find them objectionable or offensive. We also discuss hazards and harms in detail throughout the paper.
We would like to thank Zeyuan Allen-Zhu, Zexuan Zhong, Shuyan Zhou, Frank F. Xu, Qian Liu, and Ruohong Zhang for their help with the experiments and constructive feedback.
**Denoising Autoencoding.** In addition to conventional LM, the denoising autoencoding task (DAE) has also been widely used to pre-train language models [82, 24]. The inputs \(\mathbf{x}_{\setminus\hat{\mathbf{x}}}\) for DAE task are corrupted text with randomly replaced spans. Then, the language models are trained to recover the replaced tokens \(\tilde{\mathbf{x}}\).
All that remains is to show that we can uncover the full logit vector for distinct prompt queries through this API. In this section, we develop techniques for this purpose. Once we have recovered multiple complete logit vector, we can run the attack from Section 4.2 without modification.
FIRST provide a one-sentence comparison of the two summaries, explaining which \ you prefer and why. SECOND, on a new line, state only "A" or "B" to indicate your \ choice.
A multitude of previous tasks have employed structured data in the form of knowledge graphs, with knowledge graph completion representing a pervasive application. The conventional methodology (Chen et al., 2022) (Saxena et al., 2022) for completion entails defining the task as a sequence-to-sequence process, wherein incomplete triples and entities are transformed into text sequences. However, this approach is constrained by its reliance on implicit reasoning, which significantly constrains the utility of the knowledge graph itself. ReSKGC (Yu and Yang, 2023) proposes the integration of retrieval augmentation techniques with knowledge graph completion. This integration entails the selection of semantically relevant triples from the knowledge graph and their utilization as evidence to inform the generation of output through explicit reasoning. This model employs data from the Wikidata5M and WikiKG90Mv2 datasets, demonstrating superior performance compared to other existing work in a range of conditions.
The analysis by Xia et al. (2022) is the most related to ours, which examines token-level training trajectories in models of varying sizes. Our findings, however, diverge from those of Xia et al. (2022), who posit that tokens with little change in perplexity are "already learned". We identify a spectrum of token patterns, including "easy tokens" and "hard tokens" that resist convergence. Recognizing this, we propose a method of selective language modeling that targets the influential tokens, optimizing the learning process.
We present the baseline results on Chinese-LLaMA-2 and Chinese-Alpaca-2 as follows. Most of the settings are identical to those in Chinese-LLaMA.