text
stringlengths 4
4.47k
|
|---|
Broader Impact.Similar to the other LLMs, our _WizardMath_ could also generate unethical, harmful, or misleading information sometimes. Therefore, future research to address the ethical and societal implications is needed. |
Large language models (LLMs) have garnered significant attention due to their strong performance across a wide range of natural language tasks, showcasing remarkable proficiency in following instructions. Despite the rapid development of LLMs, the language capabilities of most of these models are constrained to English due to limitations in data availability, restricting the potential for application in other languages. |
An effective memory system is crucial not just for comprehending long contexts with LLMs, but also for reasoning, planning, continual adaptation for fresh knowledge, and even for learning how to learn. This work introduces a close integration of compressive memory module into the vanilla dot-product attention layer. This subtle but critical modification to the attention layer enables LLMs to process infinitely long contexts with bounded memory and computation resources. We show that our approach can naturally scale to a million length regime of input sequences, while outperforming the baselines on long-context language modeling benchmark and book summarization tasks. We also demonstrate a promising length generalization capability of our approach. 1B model that was fine-tuned on up to 5K sequence length passkey instances solved the 1M length problem. |
To enhance the reasoning ability of LLMs, [31] proposed Chain-of-Thought Prompting, which attaches multiple reasoning steps before obtaining the answer for a question. By employing the simple few-shot reasoning strategy, LLMs are able to perform better in complex reasoning problems. Least-to-Most [68] prompting decomposes the problem into sub-problems that are then solved incrementally. Additionally each step has a more detailed reasoning process. Similarly, the Complex CoT [35] underscores the pivotal role of prompt complexity by strategically choosing the most intricate problems and their corresponding solutions to function as prompts. To alleviate the burden of manual efforts, [33] introduced Auto-CoT, an approach that automates the process of acquiring k samples through the application of clustering techniques on a provided dataset. With the objective of mitigating manual intervention, [32] proposed Zero-shot-CoT, which entails the straightforward practice of appending the phrase "Let's think step by step" to each answer, eliciting the inference steps without examples. Moreover, [34] expanded upon this notion by suggesting the exploration of diverse inference paths throughout the reasoning process. Consequently, the ultimate outcome is determined through either the aggregation of answers using majority voting or by leveraging a validation mechanism, as posited by [69]. [16] employs a straightforward approach for generating augmented samples, focusing on probing the correlation between LLMs and math reasoning ability. |
Figure 4: Dataset construction pipeline for (a) UpdatedLAMA, NewLAMA, and (b) NewLAMA-Easy |
In this section, we explore X-of-thought reasoning through three different categorizations: the construction of X-of-thought (SS4.1), the structural variants of X-of-thought (SS4.2), and the enhanced methods of X-of-thought (SS4.3). |
* Continued pre-training: train on test documents without instruction-tuning (Fig. 4 ).6* Standard instruction-tuning: train on both train and test documents before instruction-tuning on train QA pairs (Fig. 4 ). |
In the main section, our EvoLLM-JP results were evolved using models found on HuggingFace. However, some of the models uses, in particular, WizardMath-7B-V1.1[33] has been released under a Non-Commercial, Research-only, Microsoft License, which is not _truly_ open-source. Therefore, our release of EvoLLM-JP is also released under a Non-Commercial, Research-only License to be consistent with the WizardMath-7B-V1.1 model. |
PersonA: While using Vascepa, you may need frequent blood tests, and it's important to follow your doctor's instructions closely. It's important to talk to your doctor about any allergies or existing conditions you have. |
**Differences in Sampling Compared to Online RL.** Standard auto-regressive sampling can bring more reasoning paths to one query. However, we find that the sampled reasoning paths have minor differences in Fig. 7, such as word substitution (orange part), rewriting of formulas (red part), rearrangement of syntax (brown part), but the correct intermediate steps are consistent (green part). |
\(\bullet\)_Memory Writing_. The purpose of memory writing is to store information about the perceived environment in memory. Storing valuable information in memory provides a foundation for retrieving informative memories in the future, enabling the agent to act more efficiently and rationally. During the memory writing process, there are two potential problems that should be carefully addressed. On one hand, it is crucial to address how to store information that is similar to existing memories (_i.e._, memory duplicated). On the other hand, it is important to consider how to remove information when the memory reaches its storage limit (_i.e._, memory overflow). In the following, we discuss these problems more in detail. (1) _Memory Duplicated_. To incorporate similar information, people have developed various methods for integrating new and previous records. For instance, in [7], the successful action sequences related to the same sub-goal are stored in a list. Once the size of the list reaches N(=5), all the sequences in it are condensed into a unified plan solution using LLMs. The original sequences in the memory are replaced with the newly generated one. Augmented LLM [43] aggregates duplicate information via count accumulation, avoiding redundant storage. (2) _Memory Overflow_. In order to write information into the memory when it is full, people design different methods to delete existing information to continue the memorizing process. For example, in ChatDB [40], memories can be explicitly deleted based on user commands. RET-LLM [42] uses a fixed-size buffer for memory, overwriting the oldest entries in a first-in-first-out (FIFO) manner. |
Determining \(B\)As long as all hyperparameters are well-tuned (especially the learning rate and regularization hyperparameters) and the number of training steps is sufficient, it is believed that the same final performance should be attainable using any batch size [19], so the batch size mainly influences the training speed of language models. Often, when training large language models, the ideal batch size is suggested to be set as the largest batch size supported by the available hardware [1], so as to maximize the training speed without considering the computational cost. In Eq 3.12, we show that the critical batch size with the optimal speed/computation trade-off can be analytically computed from the loss value. Under the guidance of this formula, we would be able to estimate the preferred batch size under any loss trajectory. Furthermore, this optimal batch size in Eq 3.12 is determined by equally minimizing the training time and required computation, as shown in Eq 3.9. In practice, if we would like to prioritize one over the other, we can follow the same process to derive the optimal batch size. By this means, we are able to obtain the optimal batch size based on our customized need in a systematic way. |
Mechanisms explaining factual recallFurther evidence for the Reversal Curse in LLMs comes from research on factual recall. Meng et al. (2023) use a model editing technique to modify factual associations. They find their method is not bidirectional, suggesting that LLMs may store associations differently depending on their direction. Complementing this, Geva et al. (2021, 2022) analyze the internal mechanisms behind factual recall in Transformers. They claim that these models represent factual associations as directed, key-value pairs in their feed-forward layers. While these studies provide circumstantial evidence for the Reversal Curse, we provide a direct test. |
Several transformer architectures have been widely adopted, with LLaMA and Mistral among the most notable. |
We use REINFORCE to optimize the likelihoods of the rationales based on their usefullness: the log-likelihood of the \(n_{true}\) true next tokens \(X_{j+1:j+n_{true}+1}\) under the language model given previous observed tokens and a particular rationale (\(p^{\text{talk}}_{j+n_{true}}\) as shorthand for the mixed prediction probabilities after thinking, see Algorithm 1). To reduce variance, we generate multiple rationale continuations for each token in the input sequence (loosely inspired by TRICE, Phan et al. (2023)). |
This paper introduces MiniCPM, comprising two SLMs with 2.4 B and 1.2 B non-embedding parameters, respectively. These models demonstrate superior performance compared to their larger counterparts. Our training methodologies are scalable both in terms of model and data size, offering potential applicability in the development of LLMs. The introduction of our WSD scheduler is notable for promoting continuous training, exhibiting compelling training dynamics, and enabling efficient study of scaling law. We further introduce the MiniCPM family, including DPO, long context and MoE versions. Future directions include in-depth analysis of the loss decrease in the decay stage, and enhancing the capability of MiniCPM by scaling in both model size and data size. |
As RALM technology develops, researchers have discovered that a single interaction is insufficient for long dialogue generation and solving multi-hop problems. Therefore, a method with multiple interactions between a retriever and a language model has been proposed. In this method, the researcher includes step s and typically has the language model generate the output first. |
Figure 3: Data distribution of the 43 programming languages used for pre-training. The number within each bar indicates the percent of the overall code distribution that an individual language comprises. |
**Overall Training Suggestion.** In practice, the above training techniques, especially 3D parallelism, are often jointly used to improve the training throughput and large model loading. For instance, researchers have incorporated 8-way data parallelism, 4-way tensor parallelism, and 12-way pipeline parallelism, enabling the training of BLOOM [78] on 384 A100 GPUs. Currently, open-source libraries like DeepSpeed [74], Colossal-AI [189], and Alpa [340] can well support the three parallel training methods. To reduce the memory redundancy, ZeRO, FSDP, and activation recomputation techniques [77, 341] can be also employed for training LLMs, which have already been integrated into DeepSpeed, PyTorch, and Megatron-LM. In addition, the mixed precision training technique such as BF16 can be also leveraged to improve the training efficiency and reduce GPU memory usage, while it requires necessary support on hardware (_e.g._, A100 GPU). Because training large models is a time-intensive process, it would be useful to forecast the model performance and detect abnormal issues at an early stage. For this purpose, GPT-4 [46] has recently introduced a new mechanism called _predictable scaling_ built on a deep learning stack, enabling the performance prediction of large models with a much smaller model, which might be quite useful for developing LLMs. In practice, one can further leverage the supporting training techniques of mainstream deep learning frameworks. |
Fuzzy deduplication.We remove similar documents by applying MinHash (Broder, 1997): for each document, we compute a sketch and measure its approximate similarity with other documents, eventually removing pairs with high overlap. |
Related work.Two works published during the responsible disclosure period use a similar procedure, and deal with numerical issues in different ways. (Chiu, 2024) start with a low \(B\) for the whole vocabulary, then increase \(B\) and ask for all tokens that haven't appeared before, and repeat until all tokens are covered. (Hayase et al., 2024) use the method in Appendix E.1, and set \(B=-\hat{z}_{i}\), where \(\hat{z}_{i}\) is an estimate of \(z_{i}\) inherent to their application. It is possible variants of this method have been discussed before our or these works, but we are not aware of further references. |
**APL-Based Logits Extraction.** Here we examine our ability to recover the next-token probability distribution. We simulate an API with a 'logit bias' argument and argmax outputs using a local LLAMA-2 model. We visualize results of our technique (blue) vs. a naive Monte Carlo sample baseline (red) in Figure 4 (left). Our deterministic algorithm extracts useful logits in fewer queries than the Monte Carlo baseline. This result follows our hypothesis (Section 3.1) that useful information is contained in the probabilities of very unlikely words, which almost never occur during random sampling. |
**Finding.** Challenging existing beliefs on data quality and LLMs, models trained on adequately filtered and deduplicated web data _alone_ can match the performance of models trained on curated data. |
We consider three realistic continual pre-training settings in the main body and provide results for a third which we believe is less warranted in the appendix. Each setting was carefully selected to expose different challenges and strengths of continual pre-training. |
Suppose queries \(b^{1},\cdots,b^{m}\) were asked and we received \(m\) answers \((i_{1},a_{i_{1}}(z,b^{1}))\leftarrow\mathcal{O}(p,b^{1}),\ldots,(i_{m},a_{i_{m}}(z,b^{m}))\leftarrow\mathcal{O}(p,b^{m})\). |
Figure 1: A motivating example of how an agent (i.e., Travel Agent) requires both LLM level and OS level resources and functions to complete a task. |
Task: Detect if the Reddit thread contains hate speech. Class label: Hate Speech Thread: All people of color are stupid and should not be allowed to vote. Class label: Not Hate Speech Thread: The best way to cook a steak on the grill. |
In the future, the advancement of RALMs will depend on the enhancement of their robustness, the improvement of retrieval quality, and the expansion of their application scope. By incorporating more sophisticated techniques and integrating RALMs with other AI technologies, these models can be leveraged to address an even broader spectrum of challenges. The ongoing research and development in this field are expected to result in more resilient, efficient, and versatile RALMs, thereby pushing the boundaries of what is achievable in NLP and beyond. As RALMs continue to evolve, they hold the promise of enabling AI systems with deeper understanding and more human-like language capabilities, thereby opening up new possibilities in a wide range of fields. |
**Output**[[Retrieve=Yes] <p>Reference [1]<p>[ [IsRET] =Relevant] In 321 AD, the Roman Emperor Constantine the Great made Sunday a day of rest from labor, stating: "On the venerable Day of the Sun let the magistrable and people residing in cities rest, and let all workshops be closed".[ [IsRET]=Fully Supported] |
In this new dataset, the preferred translation (\(y_{w}\)) remains the best of the three translation candidates, selected in the same manner as in Section 3.1. However, the dis-preferred translation is intentionally modified to be a noised version of \(y_{w}\). We applied random deletions of words with a probability of 0.15 and word swaps within a range of 1 with a probability of 0.3, following the method suggested by Zeng et al. (2023) for creating manually noised dis-preferred data. This approach produces worse translations that are artificial. |
included in the original training corpus is supported by the low QA performance in Tab. 1 (9.5%/17.2% for 7B/70B).4 To accelerate the training process, we only use the first section of each article, which offers a thorough summary and contains many factual statements. The number of collected documents and an example document about "Oppenheimer" can be found in Fig. 2 and Fig. 3. We refer to this as the Wiki2023 dataset. |
The first two subsets are illustrated in Figure 3. |
Utilizing one of the most advanced open-source models, Llama-2 7B, our research investigates the effects on LLMs when they undergo finetuning8 with data generated by other LLMs. To ensure the generation of high-quality data and to provide a relevant but not trivial downstream task, we employ the Wikitext-103 dataset. We segment this dataset into chunks of 128 tokens, between each with a stride of 64 tokens, resulting in approximately 2.2 million chunks. Denote this dataset as \(\mathcal{D}_{0}\). The task for generation involves producing the final 32 tokens given the initial 96 tokens from each chunk in the original dataset. In the initial generation (0-th generation), we use the Llama-2 7B FT model, which has been finetuned on \(\mathcal{D}_{0}\), applying a generation loss that focuses solely on the cross-entropy loss of the final 32 tokens. We denote this initial model as \(\mathcal{M}_{0}\), which demonstrates enhanced capacity for the generation task compared to the standard Llama-2 7B model. By querying \(\mathcal{M}_{0}\) with the original 96 tokens from \(\mathcal{D}_{0}\), we generate the dataset \(\mathcal{D}_{1}\) and subsequently finetune Llama-2 7B on this dataset to obtain \(\mathcal{M}_{1}\). This process is sequentially repeated to generate \(\mathcal{D}_{i}\) from \(\mathcal{M}_{i-1}\) and obtain \(\mathcal{M}_{i}\) through finetuning. By comparing the performance of various \(\mathcal{M}\) models on the test set derived from Wikitext-103, also segmented into 128-token chunks, we aim to investigate the model collapse in LLMs. |
RAG Lewis et al. (2020) uses the FEVER dataset to map labels (Supported, Refuted, NotE-noughInfo) to individual output tokens. It is trained directly using the declaration class, which is not supervised over the retrieved evidence, unlike other works. Atlas Izacard et al. (2022) employs few-shot learning to achieve performance comparable to previous studies in just 64-shot conditions. Furthermore, after training with the full dataset, it outperformed the best model Hoffmann et al. (2022) available at the time. FILCO Wang et al. (2023) approached the task of improving the quality of retrieved documents by using the FEVER dataset from the KILT base aggregation, which only included the supports and refutes tags. Accuracy was used as a metric. |
**Process-Supervised RLHF.** In existing literature of RLHF [376], the supervision signals for RL training can be generally classified into two distinct categories: outcome-supervision signals and process-supervision signals. The outcome-supervised RLHF employs a quantitative score to assess the quality of the whole text generated by LLMs. In contrast, process-supervised RLHF offers an evaluation of each individual component (_e.g._, sentence, word, or reasoning step) within the generated content, which can provide fine-grained supervision signals to guide the training, helping LLMs refine the undesired generation contents [376, 377]. OpenAI has proposed a fine-grained annotation dataset named PRM800k [377] consisting of 12K process-annotated mathematical problems (_i.e._, MATH dataset [378]) and 75K solutions generated by LLMs of these problems, where each reasoning step of mathematical problems is labeled as _positive_, _negative_ or _neutral_ in PRM800k. This fine-grained dataset has been utilized in existing work [377, 379] to train the process-supervised reward models (PRM), and the probability from the prediction of each label can be considered as the supervision signals during RLHF procedure. To effectively leverage process-supervision signals from PRMs, existing work [376] has utilized expert iteration [380, 381], an effective RL algorithm to improve the base policy via learning from expert policy. Typically, expert iteration contains two main stages: policy improvement and distillation [376]. In the policy improvement stage, expert policy processes the systematic search procedure to produce the samples. PRMs provide process-supervision signals to guide expert policy in the search procedure and enhance the quality of samples. Subsequently, during the distillation stage, the samples generated by expert policy in the first stage are utilized to improve the base policy through supervised fine-tuning. In addition to expert iteration, PRMs can also be utilized to re-rank the candidates of the final answers generated by LLMs [377] or to select better intermediate reasoning steps during step by step reasoning [382, 379]. |
We present Nemotron-4 15B, a decoder-only transformer-based large language model. It is trained on 8 trillion tokens spanning English, 53 additional natural languages as well as 43 programming languages. Nemotron-4 15B exhibits the strongest multilingual performance of any general purpose language model at its scale - even outperforming models specialized for the multilingual domain. Nemotron-4 demonstrates that pre-training sets for large language models can continue to be scaled up even further in order to improve the abilities of models. |
Large language models possess vast amounts of knowledge on various topics. In this work, we tested their capability to adapt to new knowledge: both specialized and completely unseen. This is among the first studies to compare two prominent approaches in this domain, namely fine-tuning and retrieval augmented generation. While fine-tuning can be useful for many use-cases, we found that RAG is a more reliable choice for knowledge injection. |
To evaluate the model's legal abilities, we use \(3\) benchmarks (i) we compare the perplexity of the backbones on \(5\) types of legal documents, (ii) we enhance LegalBench with LegalBench-Instruct for deeper evaluation, (iii) we rely on the legal section of MMLU for additional insights. |
**Action Goal**: The agent can perform actions with various objectives. Here, we present several representative examples: (1) _Task Completion_. In this scenario, the agent's actions are aimed at accomplishing specific tasks, such as crafting an iron pick-axe in Minecraft [38] or completing a function in software development [18]. These actions usually have well-defined objectives, and each action contributes to the completion of the final task. Actions aimed at this type of goal are very common in existing literature. (2) _Communication_. In this case, the actions are taken to communicate with the other agents or real humans for sharing information or collaboration. For example, the agents in ChatDev [18] may communicate with each other to collectively accomplish software development tasks. In Inner Monologue [61], the agent actively engages in communication with humans and adjusts its action strategies based on human feedback. (3) _Environment Exploration_. In this example, the agent aims to explore unfamiliar environments to expand its perception and strike a balance between exploring and exploiting. For instance, the agent in Voyager [38] may explore unknown skills in their task completion process, and continually refine the skill execution code based on environment feedback through trial and error. |
To confirm our observations on LegalBench-Instruct, we analyze the results on Legal-MMLU shown in Figure 6. Again, SaulLM-7B-Instruct exhibits consistent superiority over non-legal instruction-tuned models, with a gap between \(3\) and \(4\) absolute points to the best 7B open-source competitor across the three tasks, providing additional evidence that SaulM-7B-Instruct is as a strong foundation to build models tailored to legal workflows. |
Scheduled Sampling: We utilize the Scheduled Sampling technique for Transformers Mihaylova and Martins (2019), which entails two forward passes. |
Our grading system can easily be expanded to account for new levels, each of which could have their own aggregation function. This is important as we anticipate that in the future we will need to introduce aggregation functions for multiple tests per hazard category, as well as multiple tests per persona. |
Schoenegger and Park (2023); Abolhasemi et al. (2023) evaluated GPT-4 and other LLMs on forecasting tournaments and found that they underperform the human crowd. This observation is in line with ours in Section 3.4. Unlike us, they make little or no efforts to improve these LMs on forecasting. |
**Parameter 2** (Figure 1(b)).: In the 100-exposure setting for GPT2 models on \(\mathsf{bioS}(N)\) data: |
Large language models (LLMs) memorize a vast amount of factual knowledge, exhibiting strong performance across diverse tasks and domains. However, it has been observed that the performance diminishes when dealing with less-popular or low-frequency concepts and entities, for example in domain specific applications. The two prominent approaches to enhance the performance of LLMs on low-frequent topics are: Retrieval Augmented Generation (RAG) and fine-tuning (FT) over synthetic data. This paper explores and evaluates the impact of RAG and FT on customizing LLMs in handling low-frequency entities on question answering task. Our findings indicate that FT significantly boosts the performance across entities of varying popularity, especially in the most and least popular groups, while RAG surpasses other methods. Additionally, the success of both RAG and FT approaches is amplified by advancements in retrieval and data augmentation techniques. The code and data is available at [https://github.com/informagi/RAGvsFT](https://github.com/informagi/RAGvsFT). |
PaLM-2 [75] is a more compute-efficient LLM with better multilingual and reasoning capabilities, compared to its predecessor PaLM. PaLM-2 is trained using a mixture of objectives. Through extensive evaluations on English, multilingual, and reasoning tasks, PaLM-2 significantly improves the model performance on downstream tasks across different model sizes, while simultaneously exhibiting faster and more efficient inference than PaLM. |
Comparison against baselines without retrieval.Table 2 (top) presents the baselines without retrieval. Our Self-Rag (bottom two rows) demonstrates a substantial performance advantage over supervised fine-tuned LLMs in all tasks and even outperforms ChatGPT in PubHealth, PopQA, biography generations, and ASQA (Rouge and MAUVE). Our approach also significantly outperforms a concurrent method that employs sophisticated prompt engineering; specifically, on the bio generation task, our 7B and 13B models outperform the concurrent CoVE (Dhuliawala et al., 2023), which iteratively prompts Llama2658 to refine output. |
Footnote 49: Note that we don’t aim to cover all the related research directions or domains, but instead demonstrating the use or impact of LLMs via these selected examples. |
**Context Snapshot and Restoration.** Consider that the scheduler algorithms may involve time quantum operations (e.g., Round-Robin) and agent requests may be suspended by the scheduler. This suspension happens even if the response has not been fully generated yet by the LLM. Therefore, it necessitates a mechanism to preserve the state of the LLM's generation process, ensuring that it can be accurately resumed once resources are available again. |
Unlike the other tasks, where the model has been exposed to aspects related to the topic during pretraining, _current events_ includes new information. In this case, standard regular fine-tuning not only did not improve the performance of Llama2 but also significantly degraded it. To improve the fine-tuning results, we explored augmentation of the data using paraphrases. |
Fig. 10: (a). Comparison between RNN/SSM for Event stream processing and their SSM-VIT block structure [158]; (b). |
First, we find that using more contexts as inputs do not always improve the results. Utilizing top-5 contexts as input yields better results compared to using either top-3 or top-10 contexts. Intuitively, more contexts have higher probability to contain correct answers (better recall scores). As a result, using top-5 contexts achieves better results than using top-3. However, as the number of contexts further increases, the model may suffer from "lost in the middle" phenomenon (Liu et al., 2023) and the difficulty of extracting answers from the provided context could also increase, which leads to inferior results by using top-10 contexts. |
* _Version: v13 (major update on November 23, 2023)._ _GitHub link: [https://github.com/RICAIBox/LMSurvey_Chinese](https://github.com/RICAIBox/LMSurvey_Chinese) version link: [https://github.com/RICAIBox/LMSurvey/blob/unit/assets/LLM_Survey_Chinese.pdf_](https://github.com/RICAIBox/LMSurvey/blob/unit/assets/LLM_Survey_Chinese.pdf_)* _K. Zhou and J. Li contributed equally to this work._ _The authors are mainly with Gaoling School of Artificial Intelligence and School of Information, Renmin University of China, Beijing, China; Jian-Yun Nie is with DIRO, Universite de Montreal, Canada. |
I want you act as a Prompt Rewriter. Your objective is to rewrite a given prompt into a more complex version to make those famous AI systems (e.g., ChatGPT and GPT4) a bit harder to handle. But the rewritten prompt must be reasonable and must be understood and responded by humans. Your rewriting cannot omit the non-text parts such as the table and code in #Given Prompt#:. Also, please do not omit the input in #Given Prompt#. You SHOUTLD complicate the given prompt using the following method: If #Given Prompt# contains inquiries about certain issues, the depth and breadth of the inquiry can be increased. or You should try your best not to make the #Rewritten Prompt# become verbose, #Rewritten Prompt# can only add 10 to 20 words into #Given Prompt#. |
Further, an empirical study [399] has been conducted to examine the effect of different tuning methods on language models. They compare four efficient tuning methods including serial adapter tuning [398], parallel adapter tuning [400, 410], and LoRA [145], on three open-source LLMs, namely GPT-J (6B), BLOOM (7.1B) and LLaMA (7B), for evaluation. Based on the experimental results on six math reasoning datasets, they show that these efficient-tuning methods under-perform the reference baseline GPT-3.5 on difficult tasks, while achieving a comparable performance on simple tasks. Overall, LoRA performs relatively well among these comparison methods, using significantly fewer trainable parameters. |
Footnote 14: Note: we believe Grosse et al. (2023) provide convincing justification for the approximations. |
Since Adrien's total salary was 30 percent higher than Lylah's, and Adrien earned $40000 four years ago, we can find Lylah's salary by subtracting 30% of $40000 from Adrien's salary. |
Current research on chain-of-thought theory is still in its preliminary exploration stage. |
In Section 4, we see that D4, SSL prototypes, and SemDeDup achieves significant gains on perplexity (averaged across different validation sets) and downstream accuracy (averaged across different NLP tasks) compared to baseline training. Further, we generally see that D4 outperforms SSL prototypes and SemDeDup. In this section, we provide a more fine-grained analysis of these claims across individual tasks. |
The covariance matrix \(\Sigma\) is unknown, but we can estimate it by following the bootstrapping philosophy: "the bootstrap is to the sample as the sample is to the population". We construct each bootstrap by sampling \(n=240\) times with replacement from our full dataset of 240 data points and then fitting the scaling law to this bootstrapped dataset. Repeating this for 4000 different bootstraps allows us to obtain a sampling distribution for the vector of parameters \((\log A,\log B,\log E,\alpha,\beta)\), from which we can compute a sample covariance matrix. Doing this gives us the \(\Sigma\) we need for the \(\chi^{2}\) test above, and the \(p\)-value of the resulting \(\chi^{2}\) test ends up being \(<10^{-60}\). This means that the difference between our parameters and Hoffmann et al.'s parameters is extremely statistically significant. |
**Token cost.** Recall that our attack requires that we learn the logits for several distinct prompts; and so each prompt must be at least one token long. Therefore, this attack costs at least two tokens per query (one input and one output), or a cost of \(1/2\) for each token of output. But, in practice, many models (like gpt-3.5-turbo) include a few tokens of overhead along with every single query. This increases the token cost per logit to \(\frac{2+\Delta}{4}\) where \(\Delta\) is the number of overhead tokens; for gpt-3.5-turbo we report \(\Delta=7\). |
Parallel linear time construction.We build a parallelized linear time suffix array algorithm. As a building block, we make black-box use of the SA-IS algorithm for constructing a suffix array in linear time Nong et al. (2009); Ko and Aluru (2003). Unfortunately, this algorithm is not easily parallelized directly, so we introduce a simple divide and conquer approach to parallelizing the array construction. |
* Result 5: In the 1000-exposure setting, a 2bit/param capacity ratio appears to be a **universal rule**: all models, even without MLP layers, closely achieve this ratio. * Result 6: With 100 exposures, some archs show limitations; notably, LLaMA/Mistral's capacity ratio is 1.3x lower than GPT2's, even after best-tuned learning rates. * Result 7: Further controlled experiments indicate that "gated MLP" usage leads to LLaMA/Mistral architecture's underperformance in knowledge storage. |
66. Janet sells an average of 12 fresh duck eggs daily on the farmers' market. If she sells them for $2 per egg how much does she make per week, assuming she sells at the farmers' market most every day7A: If Janet sells on average 12 fresh duck eggs daily on the farmers' market. This is equal to \(\{12\)times 7\}, so 12 \(\{\)\(\{\)times\(\(\}\)\(7\)\(7\)\(8\(\}\)\(\backslash\). The total amount she would then make would be \(\{\)84times 2\(\}\)) dollars, or 168 per week. So Janet makes $168 selling duck eggs per week. |
One notable result here is that OLMo-7B is much farther ahead of the other models on Dolma 100 Programming Languages (100 PLs). Note that this effect may be due in part to underestimation from contamination, as decontaminating code data is beyond the scope of the method in Paloma. At the same time other models that are trained on code data from GitHub such as RPJ-INCITE-7B, that are just as likely to have contamination, fair much worse. Another factor then is that OLMo-7B trains on code data with exactly the same post-processing as that in 100 PLs while the code data in other models will have been processed differently. Similarly, Pile evaluation demonstrates these in-distribution and potential contamination effects as Pythia-6.9B achieves top performance despite being trained on almost an order of magnitude fewer tokens than OLMo-7B. |
To examine the effectiveness and superiority of LLMs, a surge of tasks and benchmarks have been proposed for conducting empirical ability evaluation and analysis. In this section, we first introduce three types of basic ability evaluation of LLMs for language generation and understanding, then present several advanced ability evaluations with more complicated settings or goals, and finally discuss existing benchmarks, evaluation approaches, and empirical analysis. |
Substituting \(S_{min}\) with \(S\)Finally, after establishing the analytical relationship between \(S_{min}\) and \(S\), we can substitute \(S_{min}\) with \(S\) to link \(S\) with \(L\). |
different models, as each test example sees completely different 5 training examples across models. To overcome this randomness and make the comparisons more fair, we set aside a pool of 50 labeled data samples from the training dataset for each task, referred to as the "shot pool". For the remaining training samples, we remove their labels and utilize them as unlabeled task data, which is used in our data selection strategy utilizing task data. This particular configuration is adopted because we do not have access to unlabeled task data to evaluate the efficacy of TACP. By using this setup, we also simulate the constraints posed by scarce labeled data. Although this approach creates unlabeled task data for TACP, the size is too small, containing only 0.24 million tokens from the four tasks. |
We conduct manual inspection on the generated data to make sure no personal information or offensive contents are generated. |
12. Iris maintains a magical correspondence with a mysterious figure named Carver through enchanted typewriters. |
DPO is relatively straightforward to implement; PyTorch code for the DPO loss is provided below:import torch.nn.functional as F |
As an alternative solution to human annotation, recent work explores generating instruction-response pairs for LLM alignment by prompting them with example data or prompts and iteratively refining the results Honovich et al. (2022); Wang et al. (2022); Li et al. (2023); Xu et al. (2023). |
**Can faithfulness be evaluated automatically? (SS4)** Collecting human annotations on 26 books cost us **SS5.2K**, demonstrating the difficulty of scaling our workflow to new domains and datasets. We thus implement multiple LLM-based raters of faithfulness, following prior work such as BooookScore (Chang et al., 2023b) and FactScore (Min et al., 2023) that achieve high correlation with human judgments. However, all of our metric configurations struggle to reliably identify unfaithful claims. Our best-performing method operates similarly to "needle-in-the-haystack"-style evaluations (Kamradt, 2023; Gemini Team, 2024) by feeding as much of the book as possible into a long-context LLM along with a single claim to verify. We promote this claim-level verification task as both important for book-length summarization evaluation as well as a challenging benchmark for long-context understanding. |
Figure 6: **Hutter++ on Bigram with limited data and top-p.** The initial model is trained on \(T_{0}=100,000\) samples. It generates \(T\) samples for Gen 1. Starting from Gen 2 models are trained on data generated by the most powerful model from the previous generation. Top-p-0.95 cutting and \(\beta=3/2\). |
\(\bullet\)_Establishing self-identification for LLM._ To deploy LLMs for real-world applications, it is necessary to establish its identity and make LLMs aware of these identity information, such as name, developer and affiliation. A practical way is to create identity-related instructions for fine-tuning the LLM. It is also feasible to prefix the input with the self-identification prompt, _e.g.,_ _"The following is a conversation between a human and an AI assistant called_ChatbotName, developed by_Developer.", where ChatbotName and Developer refer to the name and developer of the chatbot, respectively. |
Figure 2: Bits per byte on 11 evaluation data sources from Paloma and their combination (Magnusson et al., 2023), decontaminated from OLMo’s pretraining data. While models follow a general data scaling trend, sample efficiency is most favorable on in-distribution data. For example, OLMo-7B overtakes all other models on C4, perhaps from having 88.8% Common Crawl pretraining data. |
Some work addresses from an empirical perspective and can serve as a practical guide. Madaan and Yazdanbakhsh (2022) decomposes prompts into three components: symbols, patterns, and text, exploring the impact of CoT through counterfactual prompting. Wang et al. (2023) analyzes the impact of demonstration selection. They find that the correctness of reasoning chains has a negligible effect, while the relevance to the question and correct reasoning order matters. Tang et al. (2023) explores the role of semantics. They find that chain-of-thought reasoning relies heavily on semantic knowledge introduced during pre-training and performs poorly in symbolic reasoning. |
Combining Real and Synthetic DataOur method of re-phrasing web data naturally incorporates the information diversity found on the internet. However, it does not incorporate the noise in real data. While synthetic data may help LLMs pre-train faster, we also want them to be able to understand noisy web text that may be filled with typos and linguistic errors so that the LLMs do not fail in user facing situations. In order to incorporate this style diversity in language modeling, we sample real and synthetic data in a 1:1 ratio. |
- McCarthy has not made a final decision and has expressed uncertainty about his future in Congress, which could mean he is still weighing his options and may not choose to resign within the given timeframe. |
We now explore the role of the infinite learning rate schedules when multiple new datasets are seen in a continual learning setup. The models are trained from random initialization with different learning rate schedules on 3 IID 100B subsets of SlimPajama (e.g., our _three datasets no shift_ setting; see Sec 5.2). We focus on the no shift setting in these preliminary experiments and leave the weak and strong shift cases to future work. This task simulates a setting where large amounts of data from the same distribution are received at time increments and we wish to continue pre-training our models on them (e.g., continuing to pre-train the model on the latest web-scrape). To make our results applicable to situations where previous optimizer states are not available, we do not keep optimizer states across dataset boundaries. Fig. 9 reports training curves for 405M parameter models. |
Any single query to \(\mathcal{O}\), no matter how well-crafted, yields at most \(\log_{2}(l)\) bits, because the output is one of \(l\) distinct values. The minimum number of queries required is at least the total information content divided by the information per query, yielding the lower bound \(l\log_{2}(B/\varepsilon)/\log_{2}(l)\). |
The benchmark has limited scopeThe v0.5 benchmark tests LMs in the English language, and is focused on the geographical and cultural context of Western Europe and North America. The benchmark only tests a single use case and three personas. |
Figure 8: Performance of Baichuan-7B, DeepSeek-7B, DeepSeek-67B, Amber-7B, OpenLLaMA-7B, and Yi-34B models across code benchmarks. |
Nonetheless, the training of such large scale models is both financially burdensome and operationally inefficient. On one hand, the empirical understanding of the mechanisms underpinning the training of LLMs remains elusive. Given the significant economic and environmental costs, experiments on LLMs are prohibitively expensive for most researchers and corporations. On the other hand, the deployment of these colossal models in everyday scenarios, such as on personal computers or smartphones, is either inefficient or unfeasible. |
Another issue of interest to us is whether the improvement in Chinese proficiency has an impact on the existing English capabilities. To address this question, we additionally collected 200,000 Chinese samples from the internet and randomly extracted 200,000 English samples from the refinedweb dataset [20]. Utilizing these samples, we evaluate the English perplexity and Chinese perplexity of LLaMA models trained on different-scale corpora, as depicted in table 2. Our findings reveal that with the increase in further pretraining scale, the perplexity of the models decreases steadily in Chinese, yet notably increases in English. This suggests that enhancing the model's capabilities solely through a single Chinese corpus comes at the cost of sacrificing the original English proficiency. |
Thus, as soon as \(T\gtrsim k^{\beta}\), the AI-generated sample size \(T\) ceases to be a "scalable" resource: collecting more AI-generated samples will not improve the performance of the downstream model, i.e performance plateaus and we lose scaling. The result is illustrated empirically in Figure 3, left and Figure 8 (Appendix B). |
Nevertheless, the evaluation of the Pythia model suite (Biderman et al., 2023) has shown that training small models on very large corpora could lead to _saturation_, in the form of a performance degradation in late pretraining. In this paper, we explore this saturation phenomenon through the lens of representation degeneration, and find that both phenomena strongly correlate. We further demonstrate that representation degeneration strongly occurs in the language modeling head of small models, and we theoretically and empirically show how a linear language modeling head can represent a performance bottleneck for architectures based on small hidden dimensions. |
To identify neurons responsible for a specific language, it is crucial to discern the significance of a neuron in relation to the inference of a given |
**Reference [1]** A Question and Answer Guide to Astronomy is a book about astronomy and cosmology, and is intended for a general audience. The book was written by Pierre-Yves Bely, Carol Christian, and Jean-Rene Roy, and published in English by Cambridge University Press in 2010. It was originally written in French. The content within the book is written using a question and answer forum. It contains some 250 questions, which The Science Teacher states each are answered with a concise and well-formulated essay that is informative and readable. The Science Teacher review goes on to state that many of the answers given in the book are little gems of science writing. The Science Teacher summarizes by stating that each question is likely to be thought of by a student, and that the answers are informative, well constructed, and thorough. The book covers information about the planets, the Earth, the Universe, practical astronomy, history, and awkward questions such as astronomy in the Bible, UFOs, and aliens. Also covered are subjects such as the Big Bang, comprehension of large numbers, and the Moon illusion. |
Chen Ma is currently pursuing a Master's degree at Renmin University of China, Beijing, China. His research interests include recommender system, agent based on large language model. |
Figure 3: Examples of mistakes in label prediction made by Claude-3-Opus and GPT-4-Turbo accompanied by annotator labels and reasoning. More examples can be found in Figure 11. |
Currently, more and more studies has begun to pay attention to the importance of data quality of instruction tuning. LIMA Zhou et al. (2023) only uses 1,000 high-quality instructions and outputs for SFT, and does not even need to perform RLHF training to achieve very strong performance. Alp-Gasus Chen et al. (2023) uses powerful LLM to automatically identify and filter low-quality data, resulting in high-quality instruction tuning data to improve performance and training speed. Humpback Li et al. (2023) filters out high-quality samples to fine-tune a more powerful LLM. |
During continual pretraining, a too-large proportion of the target data can hurt the performance of the original data. A representative mixture optimization target is to maintain the general-purpose ability (looses on the Pile) unchanged. To this end, using the fitted data mixing laws, we predict the _critical proportion_ leading to the same loss as before continual pretraining. Fig. 8 demonstrates the success of our prediction where the proportion we find results in similar performance compared to the model before continual pretraining while gaining improvement in the target domain. |
After SFT, to further improve the model, RL is applied to align the model to the specific goal. The model automatically samples responses, and a reward model scores these generated responses to distinguish their quality. For mathematical reasoning, it has been verified that rewards based on intermediate processes yield better results compared to rewards based solely on final results (Wang et al., 2023; Lightman et al., 2023). |
Figure 1: Comparison of fertility scores between mono- and multilingual tokenizers applied to (a) Non-English, multilingual documents and (b) English documents. |
The most prevalent LMs nowadays are the GPT families (Radford et al., 2019) (Brown et al., 2020) (Achiam et al., 2023) and Bert families (Liu et al., 2019) (Devlin et al., 2018)(Sanh et al., 2019), which have been demonstrated to excel in a multitude of natural language processing tasks. Among these, the AutoEncoder language model is particularly adept at natural language understanding tasks,while the AutoRegressive language model is more suited to natural language generation tasks. While increasing parameters (Touvron et al., 2023b) and model tuning (Han et al., 2023) can enhance the performance of LLMs, the phenomenon of "hallucination" (Ji et al., 2023) persists. Furthermore, the limitations of LMs in effectively handling knowledge-intensive work (Feng et al., 2023) and their inability to promptly update their knowledge (Mousavi et al., 2024) are consistently apparent. Consequently, numerous researchers (Lewis et al., 2020) (Izacard and Grave, 2020b) (Khandelwal et al., 2019) have employed the technique of retrieval to obtain external knowledge, which can assist the language model in attaining enhanced performance in a multitude of tasks. |
**Reference [1]** Ronaldo Assis Moreira (born 21 March 1980), commonly known as Ronaldinho Gaichco or simply Ronaldinho, is a Brazilian retired professional footballer who played mostly as an attacking midfielder, but was also deployed as a winger. Considered one of the best players of his generation and regarded by many as one of the greatest of all time, Ronaldinho won two FIFA World Player of the Year awards and a Ballon d’Or. A global icon of the sport, he was renowned for his technical skills, creativity, dribbling ability and accuracy from free-kicks, as well as his use of tricks, feints, no-look passes and overhead... |
Our goal is to create a unified framework capable of automatically generating a merged model from a selection of foundation models, ensuring that the performance of this merged model surpasses that of any individual in the collection. Central to our approach is the application of evolutionary algorithms, which we employ to refine the intricacies involved in model merging. To systematically address this challenge, we first dissect the merging process into two distinct, orthogonal configuration spaces, analyzing their individual impacts. Building on this analysis, we then introduce a cohesive framework that seamlessly integrates these spaces. Figure 1 provides a schematic representation of our approach. |
we have shown that logit outputs are identical and so again we cannot distinguish these transformers by querying \(\mathcal{O}\) and \(\mathcal{O}^{\prime}\) alone. |
**Are there tasks for which the dataset should not be used?** The dataset is not intended for training the models. It was created solely for the purpose of testing the models. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.