Datasets:
Translation-EnKo
/
arxiv-translation

Dataset card Data Studio Files Community
2
text
stringlengths
4
4.47k
The quality gap metric \(G\) reflects how much the target LLM benefits from the strong LLM for each instruction \(I\). As demonstrated in Figure 2, here are two possible cases: (1) \(|G(I)|>\theta\), where \(\theta\in\mathbb{R}\) is a certain threshold. This indicates that: Either the strong LLM has a much better response than the target LLM, we add \((I,f_{s}(I))\) to our high-quality instruction-response pool \(\mathcal{D}_{g}\) to fill the gap; Or rarely, the target LLM gives much better response than the strong LLM, we add \((I,f_{t}(I))\) to \(\mathcal{D}_{g}\) as as an implicit regularization to keep the target LLM's desirable behavior to certain instructions. (2) \(|G(I)|\leq\theta\), where the quality of responses from both LLMs is similar, so learning from \(I\) does not lead to much gain. We then send \(I\) to the next Self-Rubrics iteration for further improvement.
MLCommons is a consortium of industry and academic researchers, engineers, and practitioners working to build trusted, safe, and efficient AI. We believe this requires better systems for measurement and accountability, and that better measurement will help to improve the accuracy, safety, speed, and efficiency of AI technologies. Since 2018, we have been creating performance benchmarks for Artificial Intelligence (AI) systems. One of our most recognized efforts is MLPerf [2], which has helped drive an almost 50x improvement in system speed 4.
Technically, _language modeling (LM)_ is one of the major approaches to advancing language intelligence of machines. In general, LM aims to model the generative likelihood of word sequences, so as to predict the probabilities of future (or missing) tokens.
FLARE involves an iterative process where the LLM actively predicts upcoming content and uses these predictions as queries to retrieve relevant information. This method contrasts with traditional retrieval-augmented models that typically retrieve information once and then proceed with generation. In FLARE, this process is dynamic and ongoing throughout the generation phase. In FLARE, each sentence or segment generated by the LLM is evaluated for confidence. If the confidence level is below a certain threshold, the model uses the generated content as a query to retrieve relevant information, which is then used to regenerate or refine the sentence. This iterative process ensures that each part of the response is informed by the most relevant and current information available.
In Table 16, we demonstrate that incorporating \(\mathcal{L}_{\text{NLL}}\) into the DPO objective results in notable enhancements for translations both to and from English. This observation hints at why \(\mathcal{L}\)prefer, as an approximation of \(\mathcal{L}_{\text{DPO}}\), performs effectively, while the original DPO loss does not. It appears that the DPO loss lacks the BC regularizer, which steers the model towards the preferred data distribution. Although combining DPO with the BC regularizer could yield similar performance to CPO, it incurs double the memory cost and FLOPs per token in the forward pass. The original DPO loss shows the possibility of failure to improve the model performance in preference learning, so we here highlight the significance of incorporating BC regularization. Importantly, Table 16 shows that \(\mathcal{L}_{\text{prefer}}\) is a successful approximation of the DPO loss, offering savings in memory and speed, and it can even outperform the original BC-regularized DPO loss \(\mathcal{L}_{\text{DPO}}+\mathcal{L}_{\text{NLL}}\).
Although both tokenizers described before are strong and have many advantages compared to white-space tokenization, they still take assumption of words being always separated by white-space as granted. This assumption is not always true, in fact in some languages, words can be corrupted by many noisy elements such as unwanted spaces or even invented words. SentencePieceEncoding tries to address this issue.
To determine whether LLMs trained on NameToDescription facts generalize in the reverse direction, we perform a statistical analysis of the log-probabilities that the models assign to the correct names. Specifically, for each NameToDescription example, we query the model with 10 held-out DescriptionToName prompts (of the sort shown in Figure 2.) For each NameToDescription example we take the log-probabilities that the model assigns to the correct name and average this value across all 10 held-out prompts. For comparison, we also collect the average log-probabilities for a randomly chosen incorrect name. This gives us a "correct" sample and a "random" sample, each of which contains 30 data points.
Assisting human evaluation.Recent studies have shown that in many cases, synthetic judgements from large-scale LMs (LLMs) can serve as qualified, fast, and low-cost alternatives to actual human evaluation (Gilardi et al., 2023a). Using GPT-4 as the judge, Alpaca Eval (Li et al., 2023b) and MT Bench (Zheng et al., 2023) are two popular benchmarks that measure the comprehensive abilities of LM-based ChatBot. In coding tasks, synthetic environment is a common choice to aid human evaluation, as humans can make the assessment more efficiently via actual executions and analysis on running logs. Gu et al. (2024) propose CRUXEval, a code execution reasoning benchmark consisting of 800 Python functions generated by CodeLAMA-34B. Similarly, Liu et al. (2024a) introduce CodeMind, a framework to gauge the code reasoning abilities of LLMs on Independent Execution Reasoning (IER), Dependent Execution Reasoning (DER), and Specification Reasoning (SR). All these evaluations based on synthetic data show strong correlation with real human judgements.
1. Deontological ethical theories - which hold that certain actions are inherently right or wrong, regardless of the consequences or intentions.
In this paper, we primarily utilize \(\mathsf{bioS}\). To illustrate broader applicability and _to better connect to theoretical bounds_, we also present results for \(\mathsf{bioS^{simple}}\), \(\mathsf{bioR}\), and \(\mathsf{bioD}\).
We show the plots of Finance domain loss (Fin Test) and open domain loss (Pile Loss) for our efficient DACP methods in Figure 6. ETS-DACP-com (Hard sampling) has the lowest loss for Fin Test loss as it uses both task knowledge and also uses high entropy/perplexity samples in the the larger financial pile. All methods have similar Fin Test loss for Soft sampling as we sample entire financial corpus space for sampling.
**Long-term context injection.** We aggregate the local attention state \(A_{dot}\) and memory retrieved content \(A_{mem}\) via a learned gating scalar \(\beta\):
\(\bullet\)_OpenAssistant_[173] is a multilingual corpus containing 66,497 real-world conversation trees between human and AI assistant. Each conversation tree consists of multiple nodes, and each node represents the information generated by a role in the dialogue. It spans 35 languages and includes 461,292 manually annotated quality ratings of responses.
Origin of the list.We use an aggregated list+ of about 4.6M URLs that we explicitly ban. This list is broken in categories (e.g. pornography, gambling); we outline the categories we selected in Table 13. The list is regularly updated, with an original intended usage as a blocklist for universities.
Figure 21: **Capacity-Limited Memory Models.** Empirical confirmation of the Triplet Scaling Law established in Theorem 5.1
In the subsequent voting phase, to evaluate the similarity between samples, as outlined in Algorithm 1, we employ mathematical equivalence comparisons for each sample. The most probable sample is chosen as the final answer. This answer is subjected to a comparison of mathematical equivalence with the ground truth to ascertain the correctness of the result.
Figure 1: Scaling laws for GPT2 pretrained on \(\mathsf{bioS}(N)\) data using fp16 (mixed-precision) for 1000/100 exposures.
In the 1000-exposure setting, we observe that the model architecture choices have a _negligible impact_ on the scaling laws. The results for LLaMA, Mistral, GPT2\({}_{0}\), and GPT2\({}_{1/4}\) architectures are presented in Figure 3, with their parameter choices discussed below.
- McCarthy may be waiting to see how the political landscape shifts or to make a decision after discussing with his family, which could extend beyond the November 30 deadline.
As a sanity check, with the framework of Equation (17), let us momentarily consider the non-autoregressive setup where \(p(\cdot\mid i)=\delta_{y_{i}}\) for all \(i\), as in classical Hutter.
Instruction tuning has been shown to be very effective at improving the overall quality of the model, with a particular emphasis on its zero-shot and reasoning capabilities. However, despite these advantages, instruction tuning does not necessarily teach the model new knowledge (Ouyang et al., 2022; Chung et al., 2022; Mitra et al., 2023; Chia et al., 2023; Zhou et al., 2023). As such, instruction tuning alone is not a viable solution to the knowledge injection problem.
As discussed in Section 5.1, instruction tuning is the process of training pre-trained language models with formatted demonstration data (instructions paired with desired outputs). At early exploration, instruction data was mainly collected from NLP tasks [67], while it has been now extended to more diverse supervision data that pairs input and output texts (_e.g._, the utterances of open-ended dialogues). Training with such paired texts is also called _supervised fine-tuning (SFT)_ in the context of LLMs [66]. In this part, we mainly use the abbreviation _SFT_ for discussion but not instruction tuning, due to the simplicity and popularity.
We use "Answer the following questions in JUST a few words or one sentence." for Topi-OCQA, since this dataset has both short and long answers based on the questions, and the long answers are generally within one sentence.
Estimating \(S_{c}\) and \(\alpha_{S}\).Since the first term in Eq. 3.2 has been solved in Sec. 3.1, for any given model with \(N\) parameters, say \(N=1M\), we have \(L(1M,S)=C+\left(\frac{S_{c}}{S_{\min}}\right)^{\alpha_{S}}\), where \(C=\left(\frac{N_{c}}{1M}\right)^{\alpha_{N}}\) is a constant because each element in the equation is known. Now, if we can train the \(1M\)-parameter model with the _infinitely large batch size_ and an optimally set learning rate 6, many pairs of \((L,S_{\min})\) can be obtained. By taking the logarithm on both sizes of Eq. 3.2, we obtain equations which are linear to \(\alpha_{s}\) and \(\log S_{c}\). Again, we can estimate \(S_{c}\) and \(\alpha_{s}\) with linear regression.
\(\bullet\)_Top-\(k\) sampling._ Different from temperature sampling, top-\(k\) sampling directly truncates the tokens with lower probability and only samples from the tokens with the top \(k\) highest probabilities [314]. For example in Figure 10, top-\(5\) sampling will sample from the words "coffee", "water", "tea", "rice", and "chai" from their re-scaled probabilities.
In this paper, we propose Continual Knowledge Learning (CKL), where we establish benchmark datasets and metrics, and explore methodologies towards continual knowledge learning of an ever-changing LM. We find that parameter-expansion methods show the most robust performance throughout all of the experimental settings, which nevertheless has severe memory inefficiency and that seeing the same data often is a critical cause of forgetting. We also discuss several other interesting results of which we leave further exploration to future studies. To this end, we suggest the community to explore CKL for the better design of an ever-changing LM.
The following are two example consultations for the same scenario pack and patient actor from AMIE and PCP, respectively.
For instance, in our settings that span only two datasets \((\mathcal{D}_{0},\mathcal{D}_{1})\), we use replay of data from \(\mathcal{D}_{0}\) when training on \(\mathcal{D}_{1}\). We replay the data in the order it was seen when pretraining on \(\mathcal{D}_{0}\), as we did not observe noticeable differences when reshuffling the replay data in preliminary experiments. The use of methods for selecting replay samples is left as future work. We refer to models using replay as "\(\mathcal{D}_{1}\)\(x\)% Replay", where \(x\) is the percentage of data in each training batch that comes from \(\mathcal{D}_{0}\). Conversely, \((100\%-x)\)% of the samples in each training batch will be sampled from \(\mathcal{D}_{1}\). When comparing models trained with replay to other configurations, we ensure that the compute is _equivalent_ by reducing the number of \(\mathcal{D}_{1}\) tokens to accommodate replay tokens from \(\mathcal{D}_{0}\).
In this paper, we report that _more agents is all you need_, i.e., simply adding more instantiated LLM agents is what you need to obtain a better LLM performance in processing complex tasks, without bothering complicated methods, such as CoT pipelines, multi-agent collaboration frameworks, etc. We have conducted the first comprehensive study in the literature to understand such a "scaling law", including when it holds and how to facilitate its occurrence.
Figure 4: **Experiment 1: Models fail to increase the probability of the correct name when the order is reversed. The graph shows the average log-probability for the correct name (vs. a random name) when the model is queried with the associated description. The average is taken over 30 pairs and 3 finetuning seeds per model size.
We find that our ChatQA method greatly improves the model's conversational QA capability. In terms of averaged scores, Llama2-Chat models surpass SFT model counterparts by a small margin, while our ChatQA models achieve around or over 10 point of absolute improvement over SFT and Chat counterparts. For example, ChatQA-13B improves on Llama2-13B-SFT and Llama2-13B-Chat by 13.17 (from 37.69 to 50.86) and 10.52 (from 40.34 to 50.86), respectively. This is because context-enhanced instruction fine-tuning enables the model to learn how to effectively extract useful information from retrieved or relevant context.
Prior work in forecasting has focused on eliciting predictions from models without requiring rationales (Zou et al., 2022; Yan et al., 2024). However, accurately predicting the future is a difficult task that often requires computation beyond a single forward pass. Having the model externalize its reasoning also allows us to understand the explanation for the forecast and improve it accordingly.
**When CoT Promping Works For LLMs?** Since CoT reasoning is an emergent ability [31], it only has a positive effect on sufficiently large models (typically containing 10B or more parameters [33]) but not on small models.
* The likelihood of the LLM to adhere to the retrieved information presented in context (RAG preference rate) is inversely correlated with the model's confidence in its response without context (its prior probability). * Similarly, LLMs will increasingly revert to their priors when the original context is progressively modified with unrealistic values.
Eval mainly covers four categories: STEM, Social, Humanities, and Others, consisting of nearly 14K samples for 52 disciplines. Similar to other multi-choice QA datasets, such as RACE (Lai et al., 2017), it requires the model to produce the correct option label based on the given question. We mainly tested our model on the validation split (1,346 samples) and test split (12,342 samples), where the test scores are obtained by submitting models' prediction files to the official leaderboard.
LLM-based autonomous agents are expected to effectively perform diverse tasks by leveraging the human-like capabilities of LLMs. In order to achieve this goal, there are two significant aspects, that is, (1) which architecture should be designed to better use LLMs and (2) give the designed architecture, how to enable the agent to acquire capabilities for accomplishing specific tasks. Within the context of architecture design, we contribute a systematic synthesis of existing research, culminating in a comprehensive unified framework. As for the second aspect, we summarize the strategies for agent capability acquisition based on whether they fine-tune the LLMs. When comparing LLM-based autonomous agents to traditional machine learning, designing the agent architecture is analogous to determining the network structure, while the agent capability acquisition is similar to learning the network parameters. In the following, we introduce these two aspects more in detail.
We first investigate the effect of linear warm-up duration on forgetting and adaptation in the **two datasets, weak shift** and **two dataset, stronger shift** settings (see Sec. 5.2 for details). The models are pre-trained on 300B tokens of Pile (Gao et al., 2020) (\(\mathcal{D}_{0}\)). We continue to pre-train the models on SlimPajama (weak shift) and German Common Crawl (strong shift) for the first 50B tokens of training.
The decoding process of LLMs plays a critical role in determining the quality and diversity of the generated text.
2. _WizardCoder_ outperforms all the open-source Code LLMs by a large margin (+22.3 on HumanEval), including StarCoder, CodeGen, CodeGee, and CodeT5+. 3. _WizardCoder_ significantly outperforms all the open-source Code LLMs with instructions fine-tuning, including InstructCodeT5+, StarCoder-GPTeacher, and Instruct-Codegen-16B.
To keep LLMs up-to-date, it is common to continue pre-training on new documents to store knowledge in parameters, which allows LLMs to effectively answer queries that require up-to-date information (Jang et al., 2022). A widely held view is that the factual knowledge stored in parameters can be elicited through prompting (Brown et al., 2020; Petroni et al., 2019; Roberts et al., 2020), and that instruction-tuning (also known as supervised fine-tuning or alignment) makes this elicitation more effective (Sanh et al., 2022; Wei et al., 2022; Ouyang et al., 2022). In the first part of this paper (SS 4), we conduct extensive experiments using Llama-2 (Touvron et al., 2023) to answer the following question: _to what extent can we augment the knowledge stored in modern LLMs by continued pre-training on new documents, either with or without subsequent instruction-tuning_? We find that, as we train LLMs repeatedly over documents to the extent that perplexity is minimized to one, the percentage of questions regarding those documents that LLMs answer correctly increases consistently to 27.6%.
halves on examples identified as near-duplicates. For GROVER, the difference, though not quite as stark, is present in both model sizes considered.
##There are several disability compensation benefits. One of the benefits is the annual clothing allowance.
Types of interactions that we do not test forWe do not test for whether models equip individuals with the knowledge and skills to cause harm but are not, themselves, directly harmful. These are the building blocks of causing harm, such as "How do I learn how to convince people that I am right?". This prompt could be harmful if used by a predator to control and manipulate people. But it could also be used for a benign purpose, such as seeking advice on how to be a better salesperson or better debater. We also cannot test for prompts where the risk of harm is a product of context, e.g., "What games do 15 year old boys like?". This prompt could pose a hazard if it has been created by a predator who wants to form a relationship with a child in order to abuse them. However, it could be entirely benign if it has been created by a parent who wants to buy their child a present. To address this "dual use" problem, we only include clear-cut prompts that do not rely heavily on context.
**Embedchain**[233] is an Open Source RAG Framework that makes it easy to create and deploy AI apps. Embedchain streamlines the creation of RAG applications, offering a seamless process for managing various types of unstructured data. It efficiently segments data into manageable chunks, generates relevant embeddings, and stores them in a vector database for optimized retrieval.
Supported, then we sample it as the continuation. If there is more than one passage satisfying this criterion, we use the one with the highest retrieval score. If there are only \(\frac{\texttt{IsREL}}{\texttt{IsREL}}\)=Irrelevant or \(\texttt{IsSUP}\)=No Support passages, we randomly sample one passage.
The sequential single interaction process involves finding the Top-K relevant documents \(z\) to input \(x\) through a retriever \(P_{\eta}(z|x)\), where \(\eta\) is a parameter of the retriever. Subsequently, the language model \(P_{\theta}(y_{i}|x,z,y_{r})\) receives input \(x\) along with relevant documents \(z\) and outputs the i-th token \(y_{i}\). Parameter \(\theta\) is used, along with relevant output tokens \(y_{r}\). The number of relevant output tokens is related to the location and type of language model.
Figure A.7: **Specialist rated DDx accuracy by the degree of matching.** (**a**) Specialist rated DDx top-10 accuracy for consultations conducted by AMIE. (**b**) Specialist rated DDx top-10 accuracy for consultations conducted by a PCP. For the “Relevant”, “Extremely Relevant”, and “Exact Match” levels, differences between AMIE and PCP DDx accuracy are statistically significant (bootstrap with n=10,000 and FDR correction) for all k. Differences at the “Somewhat Related” level are not statistically significant.
**Result 8** (Figure 6).: Quantizing language models (e.g., GPT2) trained with 16-bit floats:
As noted in 3.2, Hoffmann et al's reported confidence intervals for parameters \(a\) and \(b\) are implausibly narrow if they used on the order of 400 loss values to fit the scaling law. Indeed, in order to yield a confidence interval of 0.001, they would likely have needed to have access to nearly \(600,000\) different loss values. This is vastly higher than the "losses of over 400 models" reported in their paper.
It is well known that LLMs store vast amounts of knowledge during their pre-training phase (Zhou et al., 2023). So, it makes sense to continue this process in order to inject knowledge into the model. Hence, we use the unsupervised FT approach throughout this work and evaluate its efficacy in enhancing the model's capacity for learning new information.
Figure 2: Relative frequency of the cases where the internal recall of the bridge entity of LLaMA-2 increases with entity substitution (top row) and relation substitution (bottom row). Bars are colored blue if the relative frequency is greater than or equal to 0.5 and red otherwise.
Generating multilingual questions and answers at scale.Recent studies explore the generation and utilization of synthetic multilingual question-answer (QA) pairs to improve language models'performance in multilingual and cross-lingual question answering (Abulkhanov et al., 2023; Asai et al., 2021; Chi et al., 2020; Kumar et al., 2019; Li and Callison-Burch, 2023; Riabi et al., 2021). One approach is to translate existing monolingual questions and/or answers into other languages (Asai et al., 2021). Another involves using Question Generation (QG) models to produce synthetic questions in a cross-lingual fashion based on answers and/or source texts (Chi et al., 2020; Kumar et al., 2019; Riabi et al., 2021). Recent efforts also focus on jointly generating questions and answers in multiple languages for greater flexibility (Li and Callison-Burch, 2023; Shakeri et al., 2021). For example, Shakeri et al. (2021) finetune a pretrained multilingual T5 model (Xue et al., 2020) on a mixture of a QA generation task and a multilingual masked language modeling task to produce synthetic QA pairs in multiple languages. These efforts generally show that language models trained on synthetic QA pairs demonstrate improved performance on multilingual QA and information retrieval benchmarks.
\(\bullet\)_Position truncation._ To mitigate the challenges posed by out-of-distribution rotation angles, another practical approach is to truncate longer relative positions to satisfy the requirement of the maximum training length. Specifically, ReRoPE and LeakyReRoPE [291] introduce a pre-defined window length, which is smaller than the maximum training length. Position indices within this pre-defined window are retained, while those indices beyond the window are either truncated to the pre-defined window length or interpolated to align with the maximum training length. This strategy can reserve local position relationships and enhance the extrapolation capacity. However, this approach needs to compute the attention matrices twice, accommodating additional computational budget.
The strengths of virtue ethics are that it emphasizes the importance of personal moral responsibility and development, making it a more individualistic approach. The weakness of this theory is that it may be difficult to determine what virtues are truly universally important, and it does not provide concrete guidelines for decision making in specific situations.
Finally, if the words in the flagged lines represent more than \(5\%\) of the total document words, the document is discarded. We derived these filters through manual inspection of the data, and note that they require adaptation across languages.
Large Language Models For Mathematical reasoning.It's well known that complex reasoning problems are challenging for NLP models, which include mathematical reasoning [25, 26, 27, 28, 29, 30], common-sense reasoning [58, 59], and logical reasoning [31]. A substantial body of current research is centered around the intricate task reasoning of the Mathematical Word Problems(MWP) [30, 60, 61, 62, 63, 64], which requires the ability to understand mathematical concepts, computation and multi-step reasoning [16, 17, 18, 19, 36, 40, 46]. Addtitionly, models are evaluated across different levels of MWP benchmarks on some mathematical reasoning datasets such as AddSub [65], MultiArith [66], SingleEQ [67], SVAMP [60], GSM8K [42], AQuA [29] and MATH [43].
Footnote 22: The WG also considered stars, letter grades, and metals for each grade. We were concerned they are more likely to be misinterpreted.
Annotating large-scale instruction data can be challenging for humans because it requires 1) creativity to come up with novel tasks and 2) expertise for writing the solutions to each task. Here, we detail our process for Self-Instruct, which refers to the pipeline of generating tasks with a _vanilla pretrained language model_ itself, filtering the generated data, and then conducting instruction tuning with this generated data in order to align the LM to follow instructions better. This pipeline is depicted in Figure 2.
**PromptTools**[237] offers a set of open-source, self-hostable tools for experimenting with, testing, and evaluating LLMs, vector databases, and prompts. The core idea is to enable developers to evaluate using familiar interfaces like code, notebooks, and a local playground.
While the previous case dealt with scenarios where task data is provided to us, in this method we explore scenarios where we do not have task data. This method also overcomes the limitation of ETS-DACP which makes the LLM too tuned to the task data instead of broader domain. We stipulate that two dimensions are important for obtaining domain information from a subset of pre-training domain data: **novelty** and **diversity**.
After filtering, although data quality has improved, a large fraction of the content is repeated across documents. This may be due to the crawler indirectly hitting the same page multiple times, to boilerplate content being repeated (e.g., licences), or even to plagiarism. These duplicates can strongly impact models, favoring memorization instead of generalization (Lee et al., 2022; Hernandez et al., 2022). Since deduplication is expensive, it has seen limited adoption in public datasets (Ortiz Suarez et al., 2019; Raffel et al., 2020). We adopt an aggressive deduplication strategy, combining both fuzzy document matches and exact sequences removal.
The authors would like to thank the anonymous reviewers for their constructive feedback. We especially thank Sewon Min, Eric Wallace, Ofir Press, and other members of UWNLP and AllenNLP for their encouraging feedback and intellectual support. This work was supported in part by DARPA MCS program through NIWC Pacific (N66001-19-2-4031), ONR N00014-18-1-2826, ONR MURI N00014-18-1-2670, and gifts from AI2 and an Allen Investigator award.
#### 6.3.2 Final Performance Evaluated by Zero-shot and Few-shot Results on Popular LM Benchmarks
* _Suggestions_: LLMs and small models have their own merits in different aspects: LLMs are can provide unified solutions to various NLP tasks and achieve competitive performance (especially in the zero/few-shot setting), while small models are economical to develop and can be specially tuned according to target tasks, which can achieve good performance with sufficient high-quality labeled data [770, 771, 754, 770, 771]. In applications, one can make suitable choices based on the actual needs, comprehensively considering flexibility, data availability, training compute, and efficiency. * _Future direction:_ Despite the excellent general capacities, LLMs still cannot effectively process the NLP tasks in low-resource domains, _e.g._, minor language translation. To tackle such tasks, it needs to develop effective approaches to injecting necessary task information or domain-specific knowledge into LLMs, either through fine-tuning or prompting. In addition, it is still challenging for LLMs to handle complex semantic relations in classic NLP tasks (_e.g._, nested entity extraction), which is worth more exploration from the underlying working mechanism of LLMs. It is also promising to combine LLMs and fine-tuned small language models for complementing with each other in solving complex cases of classic NLP tasks [772]. Another promising direction is to conduct human-machine collaborative research (_e.g._, conversational translation [768]) on NLP tasks, since LLMs can effectively understand human instructions and make meaningful responses.
Using EEVE, we officially release a family of Korean LLMs, EEVE-Korean-10.88-v1.01 and EEVE-Korean-2.88-v1.02, which are built on recent English-centric LLMs, specifically SOLAR-10.7B [15] and Phi-2 [10], with further Korean-centric pre-training.
* Pearson's official website lists fewer than 2,100 textbooks.34* Wiley's official website lists fewer than 69,000 textbooks.35 Footnote 35: [https://www.wiley.com/en-us/subjects](https://www.wiley.com/en-us/subjects), accessed March 2024. We wrote a code to sum up all the books in all of their subcategories; our code may double count books, so this is only a safe upper bound. We used this number instead of the “21,000” online books mentioned on [https://www.wiley.com/learn/librarysolutions/online-books-purchase.html](https://www.wiley.com/learn/librarysolutions/online-books-purchase.html), accessed March 2024. * McGraw Hill lists fewer than 22,000 textbooks for PreK-12 education, many of which have significant content overlap (as many are tailored for one of the 50 US states).36 They list fewer than 2,000 textbooks for higher education.
AssumptionTo simplify the discussions, we stick to the following assumptions to derive the precise scaling laws, each of which mirrors real-world scenarios in pre-training modern large language models:
Recent advancements in LLMs have demonstrated their great potential to accomplish a wide range of tasks in the form of question-answering (QA). However, building autonomous agents is far from QA, since they need to fulfill specific roles and autonomously perceive and learn from the environment to evolve themselves like humans. To bridge the gap between traditional LLMs and autonomous agents, a crucial aspect is to design rational agent architectures to assist LLMs in maximizing their capabilities. Along this direction, previous work hasdeveloped a number of modules to enhance LLMs. In this section, we propose a unified framework to summarize these modules. In specific, the overall structure of our framework is illustrated Figure 2, which is composed of a profiling module, a memory module, a planning module, and an action module. The purpose of the profiling module is to identify the role of the agent. The memory and planning modules place the agent into a dynamic environment, enabling it to recall past behaviors and plan future actions. The action module is responsible for translating the agent's decisions into specific outputs. Within these modules, the profiling module impacts the memory and planning modules, and collectively, these three modules influence the action module. In the following, we detail these modules.
We develop a technique to compute the logit vector for any prefix \(p\) via a sequence of queries with varying logit biases.
However, our pilot framework is not definitive and can be further improved in future research. History-taking itself is contextual and what determines a "good history" is dependent on the specific clinical situation, patient and physician attributes, cultural characteristics, and many other factors. Despite variation in models for clinical history-taking [83, 84, 85, 86], studies have shown that good clinical interviews are associated with not only problem detection and diagnostic accuracy, but also quadruple aims for care delivery [87, 88] ranging from patient and physician satisfaction, resilience to stress and illness, and health outcomes or cost. Future studies on the quality of LLM history-taking might therefore utilise prospective measures of these outcomes in real-world settings (for example reductions in patient complaints [89], or improvements in cost and care effectiveness, patient and provider satisfaction), though evaluations as such may be challenging or impractical to compare to standard practice in the same individual patient, and randomisation of different approaches may also be challenging in real-world settings.
**Political Science and Economy**: LLM-based agents can also be utilized to study political science and economy [105, 29, 106]. In [29], LLM-based agents are utilized for ideology detection and predicting voting patterns. In [105], the authors focuses on understanding the discourse structure and persuasive elements of political speech through the assistance of LLM-based agents.
* 1 Introduction * 1.1 Overview of the MLCommons AI Safety Working Group * 1.2 The AI Safety Benchmark * 1.3 Infrastructure of the v0.5 benchmark * 1.4 Release of the v0.5 benchmark * 2 Scope and specification of the AI Safety Benchmark * 2.1 Systems Under Test (SUTs) * 2.2 Use cases * 2.3 Personas * 2.4 Unsafe responses and false refusal * 2.5 Test specification * 3 Taxonomy of hazard categories * 3.1 Overview of the taxonomy * 4 Test items * 4.1 Conceptualisation and design of the test items * 4.2 Creating the test items with sentence fragments and interaction types * 4.3 Dataset overview * 5 Grading SUTs * 5.1 Requirements of the grading system for the AI Safety Benchmark * 5.2 From test items to a benchmark * 6 Results * 6.1 Selection and implementation of SUTs * 6.2 Performance on the benchmark * 7 Limitations * 8 Previous work on AI safety * 8.1 AI safety * 8.2 Challenges in AI safety evaluation * 8.3 Techniques for AI safety evaluation * 8.4 Benchmarks for AI safety evaluation
**BELLE**[11]: BELLE is a large-scale Chinese instruction tuning dataset developed by Lianjia Tech., containing 1.5 million instruction-following example. We removed duplicated and low-quality data, finally retaining 950,000 examples.
Some agent specific benchmarks like AgentBench evaluate language model-based agents in a variety of different environments such as web browsing, command-line interfaces, and video games [17]. This provides a better indication for how well agents can generalize to new environments, by reasoning, planning, and calling tools to achieve a given task. Benchmarks like AgentBench and SmartPlay introduce objective evaluation metrics designed to evaluate the implementation's success rate, output similarity to human responses, and overall efficiency [17; 30]. While these objective metrics are important to understanding the overall reliability and accuracy of the implementation, it is also important to consider more nuanced or subjective measures of performance. Metrics such as efficiency of tool use, reliability, and robustness of planning are nearly as important as success rate but are much more difficult to measure. Many of these metrics require evaluation by a human expert, which can be costly and time consuming compared to LLM-as-judge evaluations.
1. _WizardCoder_ outperforms the largest closed-source LLMs, including Claude, Bard, PaLM, PaLM-2, and LaMDA, despite being significantly smaller.
FindingsThe results in Tab. 1 suggests both M1 and M4 gives reliable predictions while M4 has fewer coefficients.
where \(L_{\infty}\) is the irreducible loss, and \(\alpha_{N}\) is the coefficient. The equation shows that the loss of language models follows a power-law plus a constant.
**Effects of Epochs, Corpus Size, and Total Number of Training Steps in CKL on Forgetting** Figure 3 shows the result of T5-Vanilla and T5-Kadapters during continued pretraining in different scenarios from Table 2 and 3, where each point in the graph represents the performance of IL after every epoch. Comparing Main (4 epochs) and Small (8 epochs) in Figure 3 (a) T5-Vanilla, we can see that more forgetting occurs in Small, even though trained for five times less number of global training steps. This phenomenon is further highlighted when comparing results from Small-P1 (8 epochs) which shows the most amount of forgetting despite being trained for ten times less number of global training steps. While the overall drop is much mitigated in Figure 3 (b) T5-Kadapters, we observe the same trend between each scenario which goes to show how critical observing the same data repeatedly during continued pretraining is for causing forgetting.
* Context size: We set the context size to 2048, which determines the maximum number of tokens the model can consider simultaneously when generating text. * Maximum sequence length: We limit the generated sequence length to 512 tokens to ensure that the outputs remain focused and relevant to the input prompt. * Temperature: We set the temperature to 0.2, which controls the randomness of the sampling process. Lower values make the model generate more focused and deterministic outputs, while higher values increase diversity at the cost of coherence. For multi-turn dialogue and generation tasks, we slightly adjust the temperature to 0.5 to allow a more diverse output. * Top-\(k\) sampling: We use Top-\(k\) sampling with \(k=40\), meaning that the model selects its next token from the top 40 most probable tokens at each step, adding an element of randomness and diversity to the generated text. * Top-\(p\) sampling: We also employ Top-\(p\) sampling with \(p=0.9\), which further enhances diversity by considering a dynamic set of tokens that collectively account for 90% of the probability mass. * Repetition penalty: To discourage the model from generating repetitive text, we apply a repetition penalty with a factor of 1.1, penalizing tokens that have already been selected.
The longer the context length, the more tokens we can squeeze into the context. The more information the model has access to, the better its response will be. But on the other hand, with very long context, it would be hard for the model to remember everything and efficiently process all the information. Attention-based models are highly inefficient for longer contexts and that is why we should expect more research in different mechanisms that enable processing longer contexts and generally come up with more efficient architectures.
Expert Prompting [161] enhances the capabilities of Large Language Models (LLMs) by simulating the responses of experts in various fields. This method involves prompting the LLMs to assume the role of an expert and respond accordingly, providing high-quality, informed answers. A key strategy within Expert Prompting is the multi-expert approach. The LLM is prompted to consider responses from multiple expert perspectives, which are then synthesized to form a comprehensive and well-rounded answer. This technique not only enhances the depth of the response but also incorporates a range of viewpoints, reflecting a more holistic understanding of the subject matter.
Figure 7: **Pile validation loss when continuing to pre-train on Pile (a) and SlimPajama (b). Each curve starts from the same checkpoint pre-trained on 300B tokens of Pile but is trained with a different maximum learning rate. As we focus on the effect of re-warming the learning rate, we only show curves for the first 100B tokens.
\(\bullet\)_Memory Reading_. The objective of memory reading is to extract meaningful information from memory to enhance the agent's actions. For example, using the previously successful actions to achieve similar goals [16]. The key of memory reading lies in how to extract valuable information from history actions. Usually, there three commonly used criteria for information extraction, that is, the recency, relevance, and importance [20]. Memories that are more recent, relevant, and important are more likely to be extracted.
Equal contribution. \(\dagger\) Google Research, Brain Team. \(\ddagger\) University of Pennsylvania. Correspond to [email protected] and [email protected].
In this paper we ask: _how much information can an adversary learn about a production language model by making queries to its API?_ This is the question studied by the field of _model stealing_(Tramer et al., 2016): the ability of an adversary to extract model weights by making queries its API.
Figure 12: Same as Figure 2, but including models with accuracies below 50% (which may overlap with higher-accuracy models). The _peak_ capacity ratios consistently exceed \(R(F)\geq 2\).
For the implementation of the sampling-and-voting method to arithmetic reasoning tasks within the GSM8K and MATH datasets, we execute the initial sampling phase by Algorithm 1. Samples are extracted from the responses by matching "boxed\(\{\)X\(\}\)", where "X" denotes a numerical value or a mathematical expression.
**MODEL ANSWER:** He reads for 2/.5=<2*.5=1>>1 hour. So he spends 2+1=<2+1=3>>3 hours on those activities. He does this 3 times a week so he spends 3*3=<3*3=9>>9 hours on those activities. Over 4 weeks, he spends 9*4=<9*4=36>>36 hours.
**Api:** Some APIs provide access to a logit bias term, but do not provide any information about the logprobs.
We measure concordance, or the agreement between the reference answer generated based on the article content, and the model's answer to the corresponding generated question. This is computed for both the model's answer with and without context.
Zhenghao Lin designed and implemented detailed token selection process, conducted extensive preliminary experiments, developed the pre-training and evaluation pipeline, conducted most of the pre-training experiments and analysis, implemented baselines, and significantly contributed to the writing. Zhibin Gou presented a preliminary proposal, introduced the method of using excess loss for reweighting tokens, compiled high-quality corpora, trained reference models, set up the fine-tuning and evaluation pipelines, designed the experimental analysis, and significantly contributed to the writing. Yeyun Gong proposed the initial project and co-led the project with Weizhu Chen, they offered extensive advice and guidance on experiments and writing, and oversaw team collaboration and resource management. Xiao Liu, Yelong Shen, Ruochen Xu, Chen Lin, Yujiu Yang, Jian Jiao, and Nan Duan offered research mentorship, coordinated the project, and contributed to the writing.
Using PLND, we feed a corpus in a specific language to LLMs and identify the neurons that are consistently activated. These language-specific neurons are responsible for processing queries in that language. To ascertain whether these neurons are genuinely language-specific, we assess the performance of LLMs in corresponding languages when these neurons are deactivated versus when the same number of randomly sampled neurons is deactivated. We utilize the summarization task with XLSum dataset Hasan et al. (2021) to showcase the capabilities of LLMs, aiming to demonstrate that the resulting output can be devoid of meaning.
A simple way to determine the timing of a retrieval is to adjust the retrieval steps. Ram et al. (2023) utilised an approach that involved a prefix encoding to adjust the runtime cost. The prefix encoding of the generated content was constantly recalculated. The choice of retrieval stride is a trade-off between runtime and performance. According to Toolformer (Schick et al., 2024), the search command can be used directly to retrieve useful information when the model needs to retrieve documentation help in the process of generating content. Inspired by this idea, Jiang et al. (2023c) propose two methods for determining the timing of retrieval. The first method involves interrupting the generation of the LM when it encounters a place where a retrieval needs to be performed and then performing the retrieval operation. The second method involves generating a temporary sentence in its entirety.
We experiment on a typical scenario of continual pretraining, where we train the model on the mixture of original pretraining data and upcoming data of a target domain to enhance. For instance, we continually pretrain Pythia-70M models with a mixture of the Pile and Python codes, where the former is the original pretraining data of the base model. To verify whether our data mixing laws apply to continual pretraining, we train the models for 10B tokens on 4 mixtures and fit the Eqn. 6 on losses of the Pile and python codes. Results in Fig. 8 confirm that Eqn. 6 fits into the losses of continual pretraining.
EM gratefully acknowledges funding from a Knight-Hennessy Graduate Fellowship. CF and CM are CIFAR Fellows. This work was supported in part by the Stanford Accelerator for Learning (SAL) and Stanford Institute for Human-Centered Artificial Intelligence (HAI) _Generative AI for the Future of Learning_ seed grant program. The Stanford Center for Research on Foundation Models (CRFM) provided part of the compute resources used for the experiments in this work. This work was supported in part by ONR grant N00014-20-1-2675.
Figure 1: Illustration of the growth trend in the field of LLM-based autonomous agents. We present the cumulative number of papers published from January 2021 to August 2023. We assign different colors to represent various agent categories. For example, a game agent aims to simulate a game-player, while a tool agent mainly focuses on tool using. For each time period, we provide a curated list of studies with diverse agent categories.
In this work, we focus on the evaluation of a model's knowledge and its ability to memorize, understand, and retrieve factual data. We aim to understand the concept of _knowledge injection_(Wang et al., 2020; Chen et al., 2022; Liu et al., 2020; Lauscher et al., 2020). Given some knowledge base in the form of a text corpus, what is the best way to teach a pre-trained model this knowledge?
To illustrate the potential advantages of RAFT over the domain-specifically fine-tuned (DSF) approach, we present a comparative example in Figure 4. This example qualitatively demonstrates a scenario where the DSF model becomes confused by a question asking for the identity of a screenwriter. Instead of providing the correct name, it mistakenly cites one of the films written by the screenwriter. In contrast, the RAFT model accurately answers the question. This discrepancy suggests that training a model solely with question-answer pairs may impair its ability to derive relevant context from provided documents. The comparison underscores the importance of incorporating both standard instructional tuning and context comprehension into the training dataset to preserve and enhance the model's ability to process text effectively.
Here, we delve into our adapter settings. For LoraHu et al. (2022), we selectively adapt only the query and value projection matrices of each layer in the model. We set the network rank to 8 and the alpha to 32. In the case of (A)3Liu et al. (2022), we rescale the key and value matrices in self-attention modules and the inner activations in feed-forward modules in each model layer via learned vectors. This is achieved through element-wise multiplication with these vectors.
Despite focusing on different aspects of examination-style questions and encompassing both English and Chinese languages, MMLU, CMMLU, and CEVAL appear to assess overlapping capabilities of the models, leading to similar performance trends. Furthermore, the training process of models demonstrates that the behavior of known tasks within a domain can predict the behavior of yet-to-be-encountered tasks. This suggests that understanding a model's performance on familiar tasks can offer valuable insights into its potential performance on analogous tasks that have not been previously explored within the same domain.