Datasets:

Translation-EnKo

/

arxiv-translation

like 0

Follow

Translation-EnKo 5

\(\bullet\)_Webpages._ Owing to the proliferation of the Internet, various types of data have been created, which enables LLMs to gain diverse linguistic knowledge and enhance their generalization capabilities [82, 26]. For convenient use of these data resources, a large amount of data is crawled from the web in previous work, such as CommonCrawl [163]. However, the crawled web data tends to contain both high-quality text, such as Wikipedia and low-quality text, like spam mail, thus it is important to filter and process webpages for improving the data quality.

1. We curate the largest, most recent dataset of real-world forecasting questions to date, for evaluating and optimizing automated forecasting systems. 2. We build a retrieval-augmented LM system that significantly improves upon the baseline and approaches the human crowd performance on competitive forecasting platforms. 3. We propose and apply a self-supervised fine-tuning method to improve LM's capability in reasoning about forecasting tasks.

PersonA: Vascepa is used together with a low-fat diet and other treatments to lower total cholesterol in people with severe high cholesterol.

* **Natural Questions**[179] is a QA dataset that consists of real anonymized, aggregated queries submitted to the Google search engine as questions. An annotator is presented with a question along with a Wikipedia page from the top \(5\) search results, and annotates a long answer (typically a paragraph) and a short answer (one or more entities) if present on the page, or marks null if no long/short answer is present. * **MMLU**[180] is intended to evaluate the knowledge gained in zero-shot and few-shot scenarios. That means that MMLU assesses both the general knowledge and problem-solving ability of a model. It covers 57 subjects in STEM, humanities, social sciences, and other areas. The benchmark varies in complexity, ranging from elementary to advanced professional. It is worth mentioning that the main contribution of this dataset is for multi-task language understanding, question answering, and arithmetic reasoning. * **MBPP**[181] stands for "Mostly Basic Python Problems" and provides a benchmark for evaluating the performance of models designed for code generation. The benchmark encompasses \(974\) short Python programs including a wide range of topics, including fundamental programming concepts and standard library usage, and more. Each challenge comprises a task description, a code solution, and three automated test cases. * **HumanEval**[182] is a dataset for code generation task. This dataset consists of \(164\) hand-crafted programming challenges. Each challenge is accompanied by a function signature, docstring, code body, and multiple unit tests. The main intuition behind developing this dataset is to guarantee the exclusion of its contents from training datasets for code generation models. * **APPS**[183] is designed for code generation task focusing on the Python programming language. The APPS dataset contains a collection of \(232,444\) Python programs. Each program in the dataset has an average of \(18\) lines of Python code. Additionally, APPS offers access to a repository of \(10,000\) unique programming exercises, each with text-based problem descriptions. The final aspect to highlight is that the it includes test cases. * **WikiSQL**[184] is crafted for code generation task and it has 87,726 carefully labeled pairs of SQL queries and corresponding natural language questions from Wikipedia tables. The SQL queries comprise three subsets: test sets (\(17,284\) examples), development (\(9,145\) examples), and training (\(61,297\) examples). * **TriviaQA**[185] is designed for QA task. This dataset comprises more than \(650,000\) question-answer-evidence triples. There are \(95,000\) question-answer pairs in this dataset, each authored by trivia enthusiasts and supported by an average of six independently sourced evidence documents. These documents are automatically acquired from Wikipedia or broader web search results. The dataset is categorized into two segments, including those with authentic answers from Wikipedia and web domains, and verified sets embody the accurately answered questions along with their associated documents from both Wikipedia and online.

\(\bullet\)_ShareGPT_[148] is collected from a data collection platform where users can upload their conversations with ChatGPT or GPT-4 through the ShareGPT API. Currently, this dataset consists of approximately 90,000 conversations, including real instructions or inquiries from human and responses from ChatGPT.

In this section, we present RAFT, a novel way of training LLMs for domain specific open-book exams.

However, when running the model in higher float64 precision, we find that indeed all dimensions are used, but that the smallest dozen or so singular values are much smaller than the other singular values, an observation made by concurrent work (Cancedda, 2024).

In this method, the FT process is viewed as a direct continuation of the pre-training phase. We start with a saved checkpoint of the original LLM and train it in a causal auto-regressive manner, i.e., predicting the next token. One major difference in comparison to actual pre-training is the learning rate. Usually, one would need a much lower learning rate when continuing the pre-training of the model to avoid catastrophic forgetting (Kirkpatrick et al., 2017).

* 1 Introduction * 2 Pretraining * 2.1 Data Processing * 2.2 Tokenization * 2.3 Model Architecture * 3 Finetuning * 3.1 Data Preprocessing * 3.2 Training Method * 4 Infrastructure * 5 Safety * 6 Evaluations * 6.1 Base Model Performance * 6.1.1 Main Results * 6.1.2 Discussions * 6.1.3 In-Context Learning Study * 6.2 Chat Model Performance * 6.2.1 Automatic Evaluations * 6.2.2 Human Evaluations * 7 Capability Extension * 7.1 Long Context Modeling * 7.2 Vision-Language * 7.3 Depth Upscaling * 8 Final Discussions * A Author List and Contributions

1. Pre-training on larger datasets helps improve performance, by presumably exposing the LLM to more "knowledge". For instance, the Pythia (300B) model achieves an average score of 44.6%, outperforming the smaller C4 (85B) dataset's score of 43.5%.

Long-context models overemphasize book endings:One interesting observation is that Claude-3-Opus and GPT-4-Turbo, which both have chunk sizes \(\geq\) 100K, tend to place more emphasis on the endings of the books to the detriment of the beginning.

We begin with a simpler case where \(g_{\theta}\) does not have residual connections but a fully connected (FC) final layer. In this case, for any invertible \(h\times h\) matrix \(\mathbf{S}\), we have that \(g_{\theta}\left(p\right)=\mathbf{S}_{g\theta^{\prime}}\left(p\right)\) where \(\theta^{\prime}\) is the same as \(\theta\) except that the weights of the final FC layer are pre-multiplied by \(\mathbf{S}^{-1}\). Hence, if \(g_{\theta}\) has a final FC layer, it is impossible to distinguish between the embedding projection layer \(\mathbf{W}\) acting on \(g_{\theta}\) and the embedding projection layer \(\mathbf{W}\cdot\mathbf{S}\) acting on \(g_{\theta^{\prime}}\), given access to the output of the API \(\mathcal{O}\) only.

\(\bullet\)_GPT-3_. GPT-3 [55] was released in 2020, which scaled the model parameters to an ever larger size of 175B. In the GPT-3's paper, it formally introduced the concept of in-context learning (ICL)17, which utilizes LLMs in a few-shot or zero-shot way. ICL can teach (or instruct) LLMs to understand the tasks in the form of natural language text. With ICL, the pre-training and utilization of LLMs converge to the same language modeling paradigm: pre-training predicts the following text sequence conditioned on the context, while ICL predicts the correct task solution, which can be also formatted as a text sequence, given the task description and demonstrations. GPT-3 not only demonstrates very excellent performance in a variety of NLP tasks, but also on a number of specially designed tasks that require the abilities of reasoning or domain adaptation. Although the GPT-3's paper does not explicitly discuss the emergent abilities of LLMs, we can observe large performance leap that might transcend the basic scaling law [30], _e.g.,_ larger models have significantly stronger ICL ability (illustrated in the original Figure 1.2 of the GPT-3's paper [55]). Overall, GPT-3 can be viewed as a remarkable landmark in the journey evolving from PLMs to LLMs. It has empirically proved that scaling the neural networks to a significant size can lead to a huge increase in model capacity.

scale of losses on both the overall validation data and different validation domains. While the overall loss helps optimize the overall performance, losses on different domains show model capabilities in various aspects. We suppose that this facilitates further mixture curating for the nuanced balance of model capabilities.

As the field of LLMs is moving fast, with new findings, models and techniques being published in a matter of months or weeks [7, 8, 9, 10, 11], AI researchers and practitioners often find it challenging to figure out the best recipes to build LLM-powered AI systems for their tasks. This paper gives a timely survey of the recent advances on LLMs. We hope this survey will prove a valuable and accessible resource for students, researchers and developers.

Footnote 6: For instance, suppose Hoffmann et al. conducted 500 training runs, each generating 1,000 intermediate loss values. To account for the variance both between and within these training runs, we use the intra-group correlation coefficient, denoted as \(\rho\).

Accompanied by the exponential growth in the agent quantity and complexity, there is an increasing strain on the functionalities of LLM and OS. For example, scheduling and prioritizing agent requests in limited LLM resources poses a significant challenge. Moreover, the LLM's generation process can become time-intensive when dealing with lengthy contexts, occasionally resulting in the generation being suspended by the scheduler. This raises the problem of devising a mechanism to snapshot the LLM's current generation result, thereby enabling pause/resume behavior even when the LLM has not finalized the response generation for the current request. Furthermore, once an agent has obtained the list of available calling tools, determining the optimal sequence for invoking these tools presents yet another challenge since multiple agents may need to call the same tool. Additionally, the concurrent operation of multiple agents necessitates a robust system for memory management across different agents, while also ensuring stringent enforcement of privacy and access control measures.

**500K length book summarization (BookSum).** We further scaled our approach by continuously pre-training a 8B LLM model with 8K input length for 30K steps. We then fine-tuned on a book summarization task, BookSum (Kryscinski et al., 2021) where the goal is to generate a summary of an entire book text.

The story then matches back to Jurning childhood, producing context about her children and the relationship with my.

BERT (Birectional Encoder Representations from Transformers) [24] is one of the most widely used encoder-only language models. BERT consists of three modules: (1) an embedding module that converts input text into a sequence of embedding vectors, (2) a stack of Transformer encoders that converts embedding vectors into contextual representation vectors, and (3) a fully connected layer that converts the representation vectors (at the final layer) to one-hot vectors. BERT is pre-trained uses two objectives: masked language modeling (MLM) and next sentence prediction.

Footnote 1: Note that a LLM is not necessarily more capable than a small PLM, and emergent abilities may not occur in some LLMs.

**Models.** We train 1B, 3B, and 7B parameters autoregressive decoder-only models, based on configurations and hyperparameters similar to GPT-3 (Brown et al., 2020), diverging mostly on our use of ALiBi (Press et al., 2021). We use FlashAttention (Dao et al., 2022) in a custom codebase. We train internal models on both The Pile and RefinedWeb to control for deviations caused by our pretraining setup-we found The Pile models to perform in-line with others. For small-scale and ablation studies (first half of Section 4.2; Section 4.3), we train models to optimality according to the scaling laws of Hoffmann et al. (2022): on 27B and 60B tokens respectively for our 1B and 3B parameters models. For the main experiments demonstrating our approach (Falcon-RW models in Section 4.2), we train the models to 350GT, in line with popular public models (Brown et al., 2020; Wang and Komatuszaki, 2021; Scao et al., 2022). Note that we do not compare against the recently introduced LLaMA models (Touvron et al., 2023), as the smallest of them is trained on x2.5 more compute than our largest model, preventing a meaningful comparison from being made dataset-wise. For a more in-depth overview of the models and pretraining datasets with which we compare, see Appendix F.

**Lemma 2**.: _Two reward functions from the same equivalence class induce the same optimal policy under the constrained RL problem._

**RMSNorm** Same to Llama2 model (Touvron et al., 2023) 7B serious. We normalize the input of each transformer sub-layer, the attention layer, and the feedforward layer, with RMSNorm (Zhang and Sennrich, 2019).

Footnote 20: The model’s ability to learn from junk data is negligible; each person in bioS(\(N^{\prime}\)) appears only 0.2 times during training when \(N=200k\), or 0.05 times when \(N=50k\).

Not all models report their performance on all datasets, and because of that, the number of models for which performance is reported in different tables varies.

For general-purpose language models, diversity and quality of instruction are crucial Cao et al. (2023); Zhou et al. (2023). However, in specialized domains it is crucial to incorporate task-specific and specialized prompts to enhance performance. Our instruction fine-tuning stage involves \(2\) key components: generic (ie, non-legal) and legal instructions. The former help enhance the model's understanding and following of commands, and includes data from diverse domains such as coding, mathematics, and general conversations. For the latter we employ an extensive collection of datasets tailored to the nuances of legal domains, covering legal question answering and summarization, among others. Through this metic lous fine-tuning on instructional data, our model, SaulLM-7B-Instruct, is able to grasp legal intricacies and excels in a wide range of associated tasks.

As language models continue to evolve at an accelerated pace, an increasing number of large models with high parameter counts and exceptional performance are emerging. Tuning the parameters and internal structure of these models has become increasingly difficult and inefficient, making instruction tuning more important than ever.

Due to their sparsity, MoE models often require higher learning rates compared to dense models.

In addition to the above basic evaluation tasks, LLMs also exhibit some superior abilities that require special consider tions for evaluation. In this part, we discuss several representative advanced abilities and the corresponding evaluation approaches, including human alignment, interaction with the external environment, and tool manipulation. Next, we discuss these advanced abilities in detail.

Figure 4: Best of \(N\) baseline for \(N=\{1,4,16,64,128\}\). Performance plateaus after roughly 64-128 samples.

Let \(z\in\mathbb{R}^{\ell}\) be the (unknown) logits and let \(\mathbf{N}=\sum_{i}\exp(z_{i})\) be the normalizing constant.

For Supervised Fine-Tuning (SFT), we used both Chinese and English data. The Chinese data consisted of the full set from CQIA (Bai et al., 2024) and OL-CC, as well as high-quality data sampled from COIG-PC (Zhang et al., 2023). The English data was sampled from the OpenHermesPreferences dataset (Huang et al., 2024). The total amount of Chinese data comprised 105K pairs of instruction data, with English data adjusted to different ratios based on the volume of Chinese data. The ratios were \(1:1,2:1,4:1,\) and \(8:1\), along with configurations that included only Chinese data and only English data. Each set of experiments was trained for 3 epochs, with specific experimental results shown in Table 12.

PersonA: Urinalysis can be performed in various healthcare settings, including doctors' offices, urgent care facilities, laboratories, and hospitals. You can provide a urine sample in a specimen cup, and the sample can be analyzed on-siteor sent to a laboratory for testing. PersonU: How much urine is typically needed for the test? PersonA: Usually, only small amounts of urine, ranging from 30 to 60 ml, are required for urinalysis testing. It's a non-invasive and relatively simple test that provides valuable information about your health.

**Efficiency of Information Acquisition.** Although AMIE displayed greater verbosity compared to PCPs in terms of total number of words generated in their responses during the consultation, the number of conversational turns and the number of words elicited from the patient actors were similar across both OSCE agents, as illustrated in Figure A.11. This suggests that both AMIE and PCPs acquired a similar amount of information from the patients during the encounter. To investigate how efficient AMIE or PCPs were at gathering sufficient information to formulate a correct diagnosis, we truncated the conversations at various turn counts and used AMIE to generate differential diagnoses based on these partial conversations. Figure A.12 depicts the top-3 DDx accuracy as a function of the number of turns provided to the model. The observed accuracies plateaued within the initial 10 conversational turns for both AMIE and PCPs. This suggests that both AMIE and PCPs were able to acquire the information necessary for formulating a diagnosis within the early stages of the conversation. Additionally, the comparable performance at every turn indicates that neither AMIE nor PCPs had a significant advantage in the efficiency or quality of information acquisition.

1. **SlimOrca** This subset of the FLAN collection comprises generic instructions, offering a focused resource for various tasks Mukherjee et al. (2023); Lian et al. (2023). 2. **Meta Math Question Answering Instructions** Designed for mathematical inquiry, this dataset15 presents a range of mathematical questions, facilitating research in math-based natural language processing Yu et al. (2023). Footnote 15: Accessible at meta-math/MetaMathQA 3.

The disadvantage of studying emergent abilities in the lens of pre-training loss is that the pre-training loss is affected by the tokenizer and the distribution of pre-training corpus. The values of pre-training loss of language models trained on different corpus are not directly comparable. One possible solution is to evaluate different language models on a public validation set with the normalized perplexity [34] to account for the different vocabulary sizes.

Large Language ModelsLarge language models Brown et al. (2020); Achiam et al. (2023); Touvron et al. (2023, 2023) have made significant achievements, with impressive performance on a wide range of tasks. Currently, closed-source large language models, represented by GPT Brown et al. (2020); Achiam et al. (2023), Gemini Team et al. (2023), Groke (xAI, 2023), and Claude-2 (Anthropic, 2023), are the most advanced models in terms of performance. However, open-source models, represented by LLaMA Touvron et al. (2023), LLaMA-2 Touvron et al. (2023) and Mixtral Jiang et al. (2024), have also progressed rapidly, and have even shown competitive performance with the closed-source models on some tasks. Our work, which aims to improve the performance of open-source LLMs on mathematical tasks by fine-tuning them on synthetic data.

In this section, we discuss our post-training methods to extend the Yi base model to 200K long-context, equip it with visual understanding capability, and enhance the 6B model by depth upscaling.

Figure 1: Infini-attention has an additional compressive memory with linear attention for processing infinitely long contexts. \(\{KV\}_{s-1}\) and \(\{KV\}_{s}\) are attention key and values for current and previous input segments, respectively and \(Q_{s}\) the attention queries. PE denotes position embeddings.

Logit bias XOR logprobs.Our attack is \(10\times\) cheaper when an adversary can supply both a logit bias and also view output logprobs. This suggests a natural mitigation: prohibit queries to the API that make use of _both_ logit bias and logprobs at the same time. This type of defense is common in both the security and machine learning community: for example, in 2023 OpenAI removed the ability to combine both echo and logprobs, but with either alone being allowed; this defense would behave similarly.

Unlike most studies (Yoran et al., 2023) (Nakano et al., 2021) that directly utilise commercial search engine APIs. Komeili et al. (2021) propose an alternative approach to using multiple commercial search engine APIs. They suggest using the Bing Search API to generate a list of URLs for each query. These URLs are then used as keys to look up their page content in a lookup table constructed from public crawl snapshots, which populates a set of pages for that query. In addition, the evaluation takes into account whether the URL is from the English Wikipedia. If so, the page title is extracted from the URL and the corresponding page is searched for in the Wikipedia dump.

Our findings place domain continual pre-training as a strong alternative to building domain-specific LLMs from scratch. By being smarter about data selection for continual pre-training, we can surpass vanilla continual pre-training at a fraction of the cost. Overall, our work paves the way for developing domain-specific LLMs at a reduced cost, with implications for a wide range of applications.

In this part of our discussion, we pay attention to an in-depth analysis of individual models, focusing on identifying whether the trends in their performance remain consistent across various metrics and observing the unique patterns that emerge across different benchmarks.

**Knowledge extraction.** The "2bit/param" result is not about word-by-word memorization. Even better, such knowledge is also flexibly extractable (e.g., via fine-tuning using QAs like "What is Anya Forger's birthday?") [3] and thus can be further manipulated in downstream tasks (such as comparing the birthdates of two people, or performing calculations on the retrieved knowledge, etc.) [4]. This is because our \(\mathsf{bioS}(N)\) data is knowledge-augmented: the English biographies have sufficient text diversities [3]. We also verify in Appendix A.2 that such knowledge is extractable.

This becomes a dynamic data schedule as the number of training stages approaches the infinite limit. Therefore, the successful application of our data mixing laws on continual training signifies a promising prospect for using it to design dynamic data schedules, a more comprehensive data curating paradigm.

We now compare infinite learning rate schedules to a cosine decay schedule. We first explore a simple single-dataset pre-training setup to evaluate the feasibility of the schedule for LLM pre-training. Subsequently, we explore its benefits in our _three datasets, no shift_ setting.

Agents are designed to extend language model capabilities to solve real-world challenges. Successful implementations require robust problem-solving capabilities enabling agents to perform well on novel tasks. To solve real-world problems effectively, agents require the ability to reason and plan as well as call tools that interact with an external environment. In this section we explore why reasoning, planning, and tool calling are critical to agent success.

Training with synthetic data makes evaluation decontamination harder.The use of synthetic data in model training poses significant challenges to fair evaluation. Evaluation benchmarks are often created by referring to public text sources, such as coursework websites or forums. Consequently, it is arguable that all publicly available benchmark test cases might occasionally be included in the pre-training data of LLMs (Gao et al., 2021; Hoffmann et al., 2022). The use of synthetic data exacerbates this issue rather than mitigating it. Although the community has proposed several techniques to detect such evaluation contamination, such as _min-k9 prob_(Shi et al., 2023), which checks the probabilities of \(k\) long-tail tokens, these token-level decontamination methods are inadequate whenthe model is trained with synthetic data. Synthetic data might include rephrased versions of the benchmark data (Mattern et al., 2023; Oren et al., 2023), rendering token-level decontamination ineffective. In addition to developing more advanced evaluation contamination detection techniques, we recommend that model developers invest in creating and maintaining in-house and protected evaluation benchmarks. These proprietary benchmarks should be carefully safeguarded to prevent leakage and ensure the integrity of the evaluation process.

The logical process of an AutoEncoder is that the original input (set to x) is weighted and mapped to y, which is then inversely weighted and mapped back to z. If through iterative training the loss function L(H) is minimised, i.e. z is as close to x as possible, i.e. x is perfectly reconstructed, then it can be said that forward weighting is a successful way of learning the key features of the input. AutoEncoder language models take their name from the Denoising AutoEncoder (DAE) (Vincent et al., 2008), which is used to predict tokens that are [masked] by contextual words (these [masked] words are actually noise added at the input, typical of thinking). DAE is a technique that involves adding random noise to the input layer of data. This helps to learn more robust features when using an unsupervised approach to pre-train the weights of a deep network in a hierarchical manner.

We fit the Hoffmann et al.

decreases, as the LM distribution anneals to argmax, as well as when temperature increases, as the distribution collapses to uniform. For both top-p and top-k we note that the model requires almost all of the distribution to perform well, indicating that there is significant information in the tails.

Our result raises a number of questions. Why do models suffer the Reversal Curse? Do non-auto-regressive models suffer from it as well? Do humans suffer from some form of the Reversal Curse? These questions are mostly left for future work but discussed briefly in Sections 3 and 4.

**Civil Engineering**: In civil engineering, LLM-based agents can be used to design and optimize complex structures such as buildings, bridges, dams, roads, etc. [138] proposes an interactive framework where human architects and agents collaborate to construct structures in a 3D simulation environment. The interactive agent can understand natural language instructions, place blocks, detect confusion, seek clarification, and incorporate human feedback, showing the potential for human-AI collaboration in engineering design.

Word Tokenization.The most basic tokenization approach is the splitting of sequences based on white spaces and considering each word as a token (Bengio et al., 2000).

We use next-word prediction loss for finetuning, and only compute loss on the responses, but not system and user instructions. We use AdamW optimizer with \(\beta_{1}\) set to 0.9, \(\beta_{2}\) set to 0.999, and \(\epsilon\) set to \(10^{-8}\). We use a sequence length of 4096, alongside a batch size of 64. We set training step to 300 with a constant \(1\times 10^{-5}\) learning rate, a weight decay of 0.1, gradient clipping with a maximum threshold of 1.0, and NEFTune [34] with a noise scale of 45 for Yi-34B-Chat and 5 for Yi-6B-Chat.

An improved cost-optimal attack.It is possible to generalize the above attack to improve _both_ the query cost and token cost. Instead of issuing queries to the model that reveal 4 or 5 logit values for a single generated token, we might instead hope to be able to send a multi-token query \(\begin{bmatrix}p_{0}&p_{1}&p_{2}\dots p_{n}\end{bmatrix}\) and then ask for the logprob vector for each prefix of the prompt \(\begin{bmatrix}p_{0}\end{bmatrix}\), \(\begin{bmatrix}p_{0}&p_{1}\end{bmatrix}\), \(\begin{bmatrix}p_{0}&p_{1}&p_{2}\end{bmatrix}\) etc. OpenAI's API did allow for queries of this form in the past, by providing logprobs for _prompt_ tokens as well as generated tokens by combining the logprob and echo parameters; this option has since been removed.

The attached code allows users to generate alternate versions of each dataset used for our experiments, finetune on the datasets using the OpenAI API, and evaluate finetuned models on our datasets. Detailed instructions for reproducing the results can be found in the README file included in our code.

**Fine-tuning Dataset** We create a fine-tuning dataset comprising approximately 80,000 pairs of prompts and responses in Traditional Chinese, generated by GPT-4 with self-instruct Wang et al. (2022). Additionally, we have added Chinese-English translation and summarization data from news sources. It is important to note that our dataset exclusively consists of single-turn prompt-response pairs, and does not include multi-turn dialogues.

Figure 8: The model acquires the ability to engage in step-by-step reasoning by leveraging prior premises through data-driven learning. We add regularization to language modeling, compelling the model to grasp the deductive relationship inherent between premises and conclusions.

We deliberately set min points per centroids low and max points per centroid high so that faiss does not attempt to manually balance the clusters while doing K-Means. Sorscher et al. [47] found that explicitly class-balancing is important: they introduce the "class balance score" (see Section H of Sorscher et al. [47]) which is the expectation of the quantity \(\frac{\text{size of majority class}}{\text{size of minority class}}\) over all pairs of classes. They then set a hard limit for the class balance score of 0.5, meaning that "every class has at least 50% of the images that it would have when pruning all classes equally" [47]. We consider the unsupervised-learning analog of the class-balance score, which we refer to as the "cluster balance" score. The cluster balance score is the expectation of the quantity \(\frac{\text{size of bigger cluster}}{\text{size of smaller cluster}}\) over all pairs of clusters. Across all of our data selection methods (and choices for R) we find that this value is generally equal to or bigger than \(0.5\) without any explicit intervention. For this reason, we do not explicitly cluster balance, although we note that changing how many points are sampled from each cluster (based on properties of the cluster) is very interesting future work.

In this work, we aim to quantify the tension between LLMs' internal knowledge and the retrieved information presented in RAG settings. To tease apart these two competing forces, we query LLMs to answer questions and measure the token probabilities while introducing varying perturbations to reference documents.

The parameters \(a\) and \(c\) are fitted from the data. If the loss of selected tokens \(\mathcal{L}_{s}\) is used for fitting, then \(a>0\). Conversely, if the loss of unselected tokens \(\mathcal{L}_{us}\) is used for fitting, then \(a<0\). Therefore, we believe that training the model on selected tokens can effectively improve its performance on downstream tasks, while unselected tokens may have a detrimental effect on the model's performance in downstream tasks.

Self-Consistency has significant applications in fields where the veracity of information is critical. It is particularly relevant in scenarios like fact-checking, where ensuring the accuracy of information provided by AI models is essential. By employing this technique, prompt engineers can enhance the trustworthiness of LLMs, making them more reliable for tasks that require high levels of factual accuracy.

**llama2 \(\rightarrow\) CP \(+\) chat vector** Continual pretraining LLaMA2 on the target language corpus and then adding chat vector.

**Kernel Layer.** The kernel layer is divided into two primary components: the OS Kernel and the LLM Kernel, each serving the unique requirements of non-LLM and LLM-specific operations, respectively. This distinction allows the LLM kernel to focus on LLM specific tasks such as context management and agent scheduling, which are essential for handling LLM-related activities and are not typically within the purview of standard OS kernel functions. Our work primarily concentrates on enhancing the LLM kernel without making significant alterations to the existing OS kernel structure. The LLM kernel is equipped with several key modules, including the LLM system call interface, agent scheduler, context manager, memory manager, storage manager, tool manager, and access manager. These components are designed to address the diverse execution needs of agent applications, ensuring efficient management and execution within the AIOS framework. The specifics of these modules will be further detailed in Section 4.

where \(\left(\frac{N_{c}}{N}\right)^{\alpha_{N}}\) is Eq. 3.1, and \(S_{c}\) and \(\alpha_{S}\) are constant scalars to be estimated.

**Mixing different training distributions:** While we chose one data distribution to both select data and train on, modern LLM setups usually mix different data sources. Our method is likely complimentary to such pipelines: practitioners may use D4 to diversify and de-duplicate individual data sources and then mix data sources to provide additional diversity in their training dataset. We leave exploring the efficacy of D4 on a mix of training distributions as future work, but expect that this will yield further gains by reducing redundancy across datasets as well as within datasets.

pervised, and reinforcement learning (RL) based methods. We proceed by briefly reviewing these methods and their relation to the problem of knowledge injection.

For the training of the model variants, we utilized two distinct codebases: Axolot7 for the initial pre-training phase and LLaMA-Factory8(hiyouga, 2023) for subsequent fine-tuning. These codebases provided a strong and reliable base for our training process.

In Figure 6 of [37], they evaluate 4 tasks in BIG-Bench with the Brier Score metric and find that the emergent abilities disapper. We hypothesis that they normalize the Brier Score with the number of options in each question, otherwise the Brier Score of 0.25 on the swahili_english_proverbs task is too low for the smallest model. Four tasks have 2, 2, 4, 5 options in each question.

For the other Chinese LLaMA models (basic version), we utilize a 20GB general Chinese corpus for pre-training, which is consistent with the corpora used by Chinese BERT-wvm (Cui et al., 2021), MacBERT (Cui et al., 2020), LERT (Cui et al., 2022), and others. We also provide "Plus" version, which further expands the pre-training data to 120GB, incorporating additional data from CommonCrawl (CC) and encyclopedia sources, enhancing the model's understanding of fundamental concepts. We concatenated all the datasets and generated chunks of block size 512 for pre-training purposes.

2020; Wu et al., 2021; Xiao et al., 2021; Li et al., 2022; Gupta and Gupta, 2022) is more challenging as it introduces additional temporal information compared to visual multi-modal reasoning.

**1.

**Raw data.** We source forecasting questions from the 5 above-mentioned platforms. This yields a total of 48,754 questions and 7,174,607 user forecasts spanning from 2015 to 2024. The dataset includes 33,664 binary questions, 9,725 multiple-choice questions, 4,019 numerical questions, and 1,346 questions of other types. The questions cover a wide range of topics across the globe (Figure 10).

**Metrics.** Considering the limitations of 2-billion parameter models, our evaluation criteria go beyond just the accuracy of responses. We additionally consider factors such as usefulness, relevance, accuracy, depth, creativity, and the level of detail in the model's answers. This comprehensive method allows for a detailed evaluation of the model's response quality. Specifically, We use GPT-4 (Achiam et al., 2023) to score responses from tested LLMs in specific problem contexts, with the scoring prompt available in the Appendix. C.2. We translate the score assignment prompt template from Zheng et al. (2024).

Figure 6: **Test-Time Documents Varying**: We study how robust RAFT is to varying numbers of test-time documents that a retriever might provide. In NQ, we find that training with 4 documents leads to the best performance, but training with 2 documents is optimal for HotpotQA. However, across both datasets, training with all datasets consisting of _oracle_ documents hurts performance.

**Summarization prompt.** We sweep over 5 candidate prompts for summarization and evaluate the resulting Brier scores. The best summarization prompt gives a Brier score of 0.193 and the second gives 0.201. In this step of hyperparameter search, the ordering of the summaries, article count and reasoning prompt are randomly chosen for each question to avoid confounding.

We investigated the scaling laws of language models, specifically the relationship between model size and the total bits of knowledge stored. Our findings reveal a _precise, universal_ scaling law: a sufficiently-trained transformer (i.e., one whose training loss has plateau-ed) can store 2 bits of knowledge per parameter, even when quantized to int8, which is only 1/4 away from the information-theoretical maximum. We also examined how these scaling laws are influenced by various hyperparameters, including training duration, model architectures, floating-point precision, sparsity constraints like MoE, and data signal-noise ratios.

GPT2 with Smaller MLP.Mistral has a larger MLP layer, and it is often believed that the MLP layer serves primarily for storing knowledge, in contrast to the Attention layer. But is this truly the case?

First of all, as discussed in the Ethical Considerations, our method requires a strong LLM to generate the data, so the performance of our method depends on the quality of the LLM and may inherit bias and fairness issues from it. On the other hand, CodecLM can benefit from stronger LLMs improved with advanced bias-reducing and fairness-enhancing approaches.

The deviation of the DeepSeek models' performance from the predictions of Eq.(1) could indeed stem from the adoption of a new model scale representation. By utilizing non-embedding FLOPs per token \(M\) instead of the overall model parameters \(N\), the scaling law that was originally designed or calibrated for the parameter might not accurately capture the performance nuances associated with computational complexity per token.

ART involves a systematic approach where, given a task and input, the system first identifies similar tasks from a task library. These tasks are then used as examples in the prompt, guiding the LLM on how to approach and execute the current task. This method is particularly effective when tasks require a combination of internal reasoning and external data processing or retrieval.

We introduce Self-Instruct, a method to improve the instruction-following ability of LMs via their own generation of instruction data. On experimenting with vanilla GPT3, we automatically construct a large-scale dataset of 52K instructions for diverse tasks, and finetuning GPT3 on this data leads to a 33% absolute improvement on SuperNI over the original GPT3. Furthermore, we curate a set of expert-written instructions for novel tasks. Human evaluation on this set shows that tuning GPT3 with Self-Instruct outperforms using existing public instruction datasets by a large margin and performs closely to InstructGPT\({}_{001}\). We hope Self-Instruct can serve as the first step to align pretrained LMs to follow human instructions, and future work can build on top of this data to improve instruction-following models.

**Benchmarking Auto-evaluation.** While auto-evaluation was significantly better than random guessing at aligning with specialist preferences, it was unclear if the resulting performance was sufficient. To test this, we had 139 dialogue pairs each rated by a second specialist, and we computed the rank-order agreement of the two specialists on this subset (see figure A.18). We observed that auto-evaluation was about as accurate as the alternative specialists in predicting the first specialist's rank-order, suggesting that it is useful to leverage auto-evaluation for these criteria.

Model merging [15; 28], a recent development in the large language model (LLM) community, presents a novel paradigm shift. By strategically combining multiple LLMs into a single architecture, this exciting development has captured the attention of researchers due to its key advantage: it requires no additional training, making it an incredibly cost-effective approach for developing new models. This accessibility has fueled a surge in interest and experimentation with model merging. The Open LLM Leaderboard [20] is now dominated by merged models, showcasing its potential for democratizing foundation model development.

We provide details about our system, described at a high-level in Section 4. We specify the hyperparameters used in our (optimized) settings. Some of them are discovered by the hyperparameter sweep (Section 5.2).

In this section, we review the architecture design of LLMs, _i.e.,_ mainstream architecture, pre-training objective, and detailed configuration. Table V presents the model cards of several representative LLMs with public details.

**Human Annotation**: This evaluation method involves human evaluators directly scoring or rankingthe outputs generated by various agents [22, 29, 105]. For example, in [20], the authors employ many annotators, and ask them to provide feedback on five key questions that directly associated with the agent capability. Similarly, [159] assess model effectiveness by having human participants rate the models on harmlessness, honesty, helpfulness, engagement, and unbiasedness, subsequently comparing these scores across different models. In [79], annotators are asked to determine whether the specifically designed models can significantly enhance the development of rules within online communities.

Note that while \(\ell(y|x)\) has the same form as the pre-training loss \(L\), they are not equal. First, the pre-training loss is an average of all the tokens in all the documents pre-trained on. According to our empirical observation, the losses of different documents are not uniform. Second, if \(x\) and similar documents do not exist in the pre-training corpus, \(\ell(y|x)\) is the generalization loss, which is often related to other factors beyond the training loss, such as the model size. For example, in computer vision, a highly over-parameterized models often improve over an under-parameterized models in test performance when both models converge on the training data [9; 4].

We evaluate GSM8k [42] and MATH benchmarks [43] by employing the following CoT [31] prompt:

evaluation. Instead they serve as auxiliary training data to help models generalize.

To add these fractions, we need to find a common denominator, which is the least common multiple (LCM) of the three denominators (-4, -2, and -3). The LCM of these denominators is -12.

Given the results from the prior section, we chose to implement the improved 4-logprob attack (Section 5.3) because it is both the most query efficient attack and also the most precise attack. Switching to a different attack algorithm would increase our total experiment cost significantly, and so we do not perform these ablation studies.

_Remark._ Upon comparing the aforementioned strategies for agent capability acquisition, we can find that the fine-tuning method improves the agent capability by adjusting model parameters, which can incorporate a large amount of task-specific knowledge, but is only suitable for open-source LLMs. The method without fine-tuning usually enhances the agent capability based on delicate prompting strategies or mechanism engineering. They can be used for both open- and closed-source LLMs. However, due to the limitation of the input context window of LLMs, they cannot incorporate too much task information. In addition, the designing spaces of the prompts and mechanisms are extremely large, which makes it not easy to find optimal solutions.

To learn an objective that fosters superior translations and rejects inferior ones, access to labeled preference data is

On dynamic data curating.Our study presents a pipeline to decide on a group of fixed mixture proportions for pretraining. More sophisticated data curating can include dynamic proportions (Albalak et al., 2023) and even a curriculum that changes data domains (Chen et al., 2024). The application of our data mixing laws in continual pretraining (Sec. 5) implies the prospect of extending our findings to these settings. On top of this, we believe that it is promising to incorporate further analysis to pursue a dynamic data mixing law.

To optimize the training process, we maintain a constant learning rate of 3e-5, and adopt a strategic approach to gradually increase the batch size from 4M tokens whenever the model's loss plateaued. This incremental adjustment of the batch size, alongside maintaining all other parameters in alignment with the established Yi-6B base model configuration, was instrumental in navigating the challenges of training at scale.

than only storing information redundantly in its parameters. Future work could strengthen such paths of traversal, ultimately leading to more parameter-efficient and controllable models. Conversely, a lack of evidence would indicate more fundamental limitations of the Transformer architecture or training. It would also have critical implications for model editing: if complex facts are recalled instead of inferred, editing only base facts will never be enough since the changes cannot propagate (Onoe et al., 2023; Zhong et al., 2023; Cohen et al., 2023).

Little is publicly known about the inner workings of today's most popular large language models, such as GPT-4, Claude 2, or Gemini. The GPT-4 technical report states it "contains no [...] details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar" (OpenAI et al., 2023). Similarly, the PaLM-2 paper states that "details of [the] model size and architecture are withheld from external publication" (Anil et al., 2023). This secrecy is often ascribed to "the competitive landscape" (because these models are expensive to train) and the "safety implications of large-scale models" (OpenAI et al., 2023) (because it is easier to attack models when more information is available). Nevertheless, while these models' weights and internal details are not publicly accessible, the models themselves are exposed via APIs.