text
stringlengths 4
4.47k
|
|---|
Raffel et al. (2020) introduces a unified framework for converting all text-based language problems into text-to-text format. The aim is to explore the potential of transfer learning techniques for natural language processing. In contrast to the original transformer, a simplified version of layer normalization is used, where activations are rescaled without additional biases. After applying layer normalization, a residual skip connection (He et al., 2016) adds the input of each subcomponent to its output. Srivastava et al. (2014) is applied to the feed-forward network, the skip connections, the attentional weights, and the inputs and outputs of the entire stack. The T5 model has been widely used as a language model by many researchers, such as Hofstatter et al. (2023), Sachan et al. (2021), and Singh et al. (2021). Additionally, Chung et al. (2022) proposed instruction tuning as an approach to improve model performance. The study focused on three aspects: the number of scaling tasks, the size of the scaled model, and the fine-tuning of chain of thought data. The results showed that larger model sizes and more fine-tuning tasks significantly improved model performance. Additionally, the study found that chain of thought(CoT) significantly improves inference level. Wang et al. (2023) used this approach to tune T5 and apply it to the RALM architecture. |
Figure 16: Scaling laws for the mixture-of-experts GPT2 models with 32 experts on the bioS(\(N\)) data. |
Specifically, prompts for the NR-LLM configuration instructed the model to answer questions without additional context, whereas prompts for the AR-LLM configuration included both the question and contextual information. In contrast, the Adapt-LLM training set was constructed following the approach outlined in Section 3.1, employing a two-step process. As a result of this process, the 74.72% of the questions in NQ are marked with the (RET) token, whereas the 87.49% questions are marked for SQuAD. |
Chain-of-thought reasoning often tends to be hallucinatory, producing incorrect reasoning steps. Errors in intermediate reasoning steps can, in turn, trigger a cascade of errors. Incorporating verification to obtain feedback and subsequently refining the reasoning process based on this feedback can be a highly effective strategy for mitigating this phenomenon, which is similar to the process of human reflection. Figure 3 depicts the overview of verification and refinement. |
query attention (GQA) (Ainslie et al., 2023) for faster inference and lower memory footprint. |
Retrieval augmented generation (RAG) is often used to fix hallucinations and provide up-to-date knowledge for large language models (LLMs). However, in cases when the LLM alone incorrectly answers a question, does providing the correct retrieved content always fix the error? Conversely, in cases where the retrieved content is incorrect, does the LLM know to ignore the wrong information, or does it recapitulate the error? To answer these questions, we systematically analyze the tug-of-war between a LLM's internal knowledge (i.e. its prior) and the retrieved information in settings when they disagree. We test GPT-4 and other LLMs on question-answering abilities across datasets with and without reference documents. As expected, providing the correct retrieved information fixes most model mistakes (94% accuracy). However, when the reference document is perturbed with increasing levels of wrong values, the LLM is more likely to recite the incorrect, modified information when its internal prior is weaker but is more resistant when its prior is stronger. Similarly, we also find that the more the modified information deviates from the model's prior, the less likely the model is to prefer it. These results highlight an underlying tension between a model's prior knowledge and the information presented in reference documents. |
In this section, we present an in-depth analysis of the distribution of perplexity, similarity, and diversity within our financial corpus. Our findings reveal that all three metrics display a highly skewed distribution. Specifically, as illustrated in the top row of Figure 3, the similarity metric demonstrates a two-modal pattern, potentially attributable to the presence of two distinct sources within our financial corpus. |
Instruction generation.A series of recent works Zhou et al. (2022); Ye et al. (2022); Singh et al. (2022); Honovich et al. (2022) generate instructions of a task given a few examples. While Self-Instruct also involves instruction generation, a major difference in our case is it is task-agnostic; we generate new tasks (instructions along with instances) from scratch. |
Several factors might contribute to this behavior. Firstly, RAG not only adds knowledge to a model but also incorporates context relevant to the question, a feature lacking in fine-tuning. Additionally, fine-tuning may impact other capabilities of the model due to a degree of catastrophic forgetting. Finally, it's plausible that unsupervised fine-tuned models might benefit from further alignment through supervised or RL-based fine-tuning, as evidenced by the vastly improved performance of Orca2 over the base Llama2. |
How do annotators perceive the task?Annotators highlighted several challenges in assessing the summaries, particularly when dealing with broad claims about themes rather than specific plot points, making it difficult to find relevant supporting evidence within the text. Abstract concepts, like emotions or thematic claims, posed significant obstacles, with some annotators struggling to locate quotations that precisely supported or refuted these claims. They also pointed out the difficulty of evaluating claims that were only partially true, which required more detailed support (see Table 4 for actual comments). |
A first fundamental step is to consider _probabilistic_ ground truth labels to replace the deterministic Hutter prediction \(i\mapsto y_{i}\) with a probability distribution \(p(j|i)\) on \(\mathbb{N}_{*}\) with power law decay (as in Equation (1)). |
Footnote 32: Again, importantly, we can do so because \(P_{N+KD+(n^{\prime}-1)K+a^{\prime}}\) depends on \(\mathcal{N},\mathcal{D}_{a}\) but they can be computed using the values of \(Q_{1},\ldots,Q_{N+KD}\). |
**Planning Content.** We will regularly include new content into this survey, to make it more self-contained and up-to-date. Here, we list several potential topics that might appear in the next major version(s): (1) more experiments with larger language models for both instruction tuning and ability evaluation; (2) more detailed prompting practice; (3) training recipe; (4) more theoretical analysis and discussion; (5) more discussions on applications. |
Figure A.14: **Auto-evaluation rated DDx accuracy for scenarios conducted at both testing locations. (a)** Auto-evaluation rated top-k DDx performance of AMIE on a set of 40 scenarios conducted in both locations. (b) Auto-evaluation rated top-k DDx performance of the PCPs on a set of 40 scenarios conducted in both locations. |
To assess the impact of the tokenizers on the model downstream performance, we trained a decoder-only transformer model of size 2.6 B for each tokenizer. We trained our models for 52.6 B tokens following the scaling laws proposed by Hoffmann et al. (2022), based on the causal language modeling training objective. The hyper-parameters are described in Table 10 in the Appendix C. |
As for the {Instruction}, we use different instructions for different datasets based on the answer types. |
To effectively evaluate the LLMs' performance on these knowledge-intensive tasks, a comprehensive auxiliary dataset was collected by scraping relevant articles per topic from Wikipedia. The rationale behind selecting Wikipedia as the primary source of knowledge is its broad coverage of relevant topics and its reliability as a repository of crowd-verified knowledge. All articles pertinent to the tasks were retrieved via the official Wikipedia API2 by identifying the relevant central page per topic. |
where \(N\) is the total number of observations and \(n\) is the number of observations per group. With \(N=500,000\) and \(n=1,000\), this calculation yields \(N_{\text{eff}}\approx 1,000\), which is still far too low to obtain the tight confidence intervals reported by Hoffmann et al. |
This pattern changes after extensive training. |
Several works have studied this problem and the predominant finding is that LMs struggle to answer questions about documents they have been trained on Wang et al. (2021); Zhu and Li (2023). It is important to note, however, that these experiments were mainly conducted using relatively small LMs such as BART, T5, or GPT-2 Wang et al. (2021); Jang et al. (2022); Hu et al. (2023), using randomly initialized transformers Zhu and Li (2023), or without instruction-tuning Ovadia et al. (2023). This makes us wonder _what are the actual limits of modern LLMs to absorb knowledge from new documents and answer questions about them using the standard continued pre-training followed by instruction-tuning recipe_. In this section, we run extensive experiments using Llama-2 7B and 70B on Wiki2023-film to test their limits. |
**Pythia:** In [95], Biderman et al. introduced Pythia, a suite of 16 LLMs all trained on public data seen in the exact same order and ranging in size from 70M to 12B parameters. We provide public access to 154 checkpoints for each one of the 16 models, alongside tools to download and reconstruct their exact training dataloaders for further study. |
_(4) Mitigation Strategies._ In Theorem 3.2 we show that mixing AI-generated data with even a small amount of clean data mitigates model collapse by introducing a grokking phenomenon. The length of the plateau is of order \(k^{\beta}/\pi\), where \(\pi\) is the proportion of training data which is from the true distribution (i.e clean data). When \(\pi=0\) (i.e only AI-generated data available), this plateau goes on forever (as in (2) and (4)). When \(\pi>0\), however small, the plateau finally halts, and the error continues to decrease a la \(T^{-c}\). This grokking phenomenon holds in the setting of _deterministic_ ground truth labels (like in the models of Hutter (2021); Cabannes et al. (2023)). For transformer models, such deterministic settings are found for instance in arithmetic tasks, and we demonstrate it empirically in our GCD transformer experiments. The grokking effect becomes attenuated in probabilistic settings, where it can lead to an S-shaped learning curve (see Figure 19). We also identify regimes where adding AI data can be beneficial and discuss ways to curate "tail" data to mitigate AI-data effects. |
* Adapt-LLM consistently outperforms typical fixed strategies for question answering, such as (i) using the IR system for all questions and (ii) relying solely on the parametric memory of the LLM. * Adapt-LLM demonstrates performance comparable to strategies that rely on popularity scores to determine when to use an IR system, even without utilizing any popularity score or similar metric. It's worth noting that popularity scores are a unique feature of the PopQA dataset, rendering them inapplicable to other open-domain question answering datasets. * When Adapt-LLM decides to retrieve additional information, the results obtained with the context are significantly better than those without it. Similarly, when Adapt-LLM directly answers questions relying on its parametric memory, it achieves high accuracies. These observations indicate that the model effectively discerns when to retrieve information and when it can answer a question without further context. * The primary bottleneck for the performance of Adapt-LLM lies in the IR system. Adapt-LLM achieves much higher performance with gold passages compared to passages retrieved by the IR system. |
Figure 5: **Final loss of 405M parameter models trained on two distribution shifts.** Figures (a) and (b) are duplicated from Fig. 6 for convenient comparison. we provided three baselines and two continually pre-trained models. The baselines (light blue, dark blue, and maroon) are trained from random initialization on 300B tokens of SlimPajama, 300B tokens of Pile, and the union of both datasets (600B tokens). The continually pre-trained models (black and violet) start from a checkpoint pre-trained on 300B tokens of Pile (dark blue curve) and use 0% and 5% replay, respectively. We observe that for both distribution shifts, the combination of re-warming the learning rate and using a small percentage of replay helps to strike a balance between forgetting and adaptation. Importantly, we note that the use of replay minimally affects downstream performance compared to the models using 0% replay. |
We train our models using the _ZeRO_ optimizer strategy (Rajbhandari et al., 2019) via PyTorch's FSDP framework (Zhao et al., 2023), which reduces memory consumption by sharding the model weights and their corresponding optimizer state across GPUs. At the 7B scale, this enables training with a micro-batch size of 4096 tokens per GPU on our hardware (see Section 3.4). For OLMo-1B and -7B models, we use a constant global batch size of approximately 4M tokens (2048 instances, each with a sequence length of 2048 tokens). For OLMo-65B model (currently training), we use a batch size warmup that starts at approximately 2M tokens (1024 instances), then doubles every 100B tokens until reaching approximately 16M tokens (8192 instances). |
In this experiment, our objective is to once again evaluate the effectiveness of the Adapt-LLM model, this time focusing on its ability to accurately determine when additional context is needed. |
There is moderate evidence of the second-hop reasoning, which does not become stronger with increasing model size.Figure 4 shows the relative frequency of the cases that increasing the bridge entity recall increases the consistency. In LLaMA-2 7B, the middle and late layers exhibit a relative frequency higher than 0.5 (random chance) with statistical significance, peaking at 0.64 in layer 30. Test result with a randomly initialized model verifies 0.5 as the randomness baseline (Figure 3(d)). |
ConfigurationKILT (Petroni et al., 2021) consists of 5 different tasks and 11 datasets: Open-Domain Question Answering (Joshi et al., 2017; Kwiatkowski et al., 2019; Fan et al., 2019; Yang et al., 2018), Fact Checking (Thorne et al., 2018), Entity Linking (Hoffart et al., 2011; Guo and Barbosa, 2018), Slot-filling (Levy et al., 2017), and Knowledgeable Open Dialogue (Dinan et al., 2019). Because each task requires a different training objective than the one used during pretraining, additional finetuning is necessary. We search for the hyperparameters such as training epochs, batch size, input size, output size, and learning rate of each individual KILT task to match the T5-base dev performance reported by Petroni et al. (2021). Using the identified configurations, we perform experiments on all of the KILT tasks with the continually pretrained models for each method as the initialization checkpoints. |
Our pipeline for data generation consists of four steps: 1) generating task instructions, 2) determining if the instruction represents a classification task, 3) instance generation with either an input-first or output-first approach, and 4) filtering low-quality data. |
Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. Boolq: Exploring the surprising difficulty of natural yes/no questions, 2019. * Clark et al. (2018) Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge, 2018. * Cobbe et al. (2021) Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems, 2021. * Computer (2023) Together Computer. Redpajama: an open dataset for training large language models, 2023. URL [https://github.com/togethercomputer/RedPajama-Data](https://github.com/togethercomputer/RedPajama-Data). * OpenCompass Contributors (2023) OpenCompass Contributors. Opencompass: A universal evaluation platform for foundation models. [https://github.com/open-compass/opencompass](https://github.com/open-compass/opencompass), 2023. * DeepSeek-Al et al. (2024) DeepSeek-Al, ; Xiao Bi, Deli Chen, Guanting Chen, Shanhuang Chen, Damai Dai, Chengqi Deng, Honghui Ding, Kai Dong, Qiushi Du, Zhe Fu, Huazuo Gao, Kaige Gao, Wenjun Gao, Ruqi Ge, Kang Guan, Daya Guo, Jianzhong Guo, Guangbo Hao, Zhewen Hao, Ying He, Wenjie Hu, Panpan Huang, Erlangen Li, Guowei Li, Jiashi Li, Yao Li, Y. K. Li, Wenteng Liang, Fangyun Lin, A. X. Liu, Bo Liu, Wen Liu, Xiaodong Liu, Xin Liu, Yiyuan Liu, Haoyu Lu, Shanghao Lu, Fuli Luo, Shirong Ma, Xiaotao Nie, Tian Pei, Yishi Piao, Junjie Qiu, Hui Qu, Tongzheng Ren, Zehui Ren, Chong Ruan, Zhangli Sha, Zhihong Shao, Junxiao Song, Xuecheng Su, Jingxiang Sun, Yaofeng Sun, Minghui Tang, Bingxuan Wang, Peiyi Wang, Shiyu Wang, Yaohui Wang, Yongji Wang, Tong Wu, Y. Wu, Xin Xie, Zhenda Xie, Ziwei Xie, Yiliang Xiong, Hanwei Xu, R. X. Xu, Yanhong Xu, Dejian Yang, Yuxiang You, Shuiping Yu, Xingkai Yu, B. Zhang, Haowei Zhang, Lecong Zhang, Liyue Zhang, Mingchuan Zhang, Minghua Zhang, Wentao Zhang, Yichao Zhang, Chenggang Zhao, Yao Zhao, Shangyan Zhou, Shunfeng Zhou, Qiaho Zhu, and Yuheng Zou. Deepseek llm: Scaling open-source language models with longtermism, 2024. * Geng and Liu (2023) Xinyang Geng and Hao Liu. Openllama: An open reproduction of lama, May 2023. URL [https://github.com/openlm-research/open_llama](https://github.com/openlm-research/open_llama). * Hendrycks et al. (2021) Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset, 2021. * Henighan et al. (2020) Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun, Tom B Brown, Prafulla Dhariwal, Scott Gray, et al. Scaling laws for autoregressive generative modeling. _arXiv preprint arXiv:2010.14701_, 2020. * Hernandez et al. (2021) Danny Hernandez, Jared Kaplan, Tom Henighan, and Sam McCandlish. Scaling laws for transfer, 2021. * Hoffmann et al. (2022) Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal large language models. _arXiv preprint arXiv:2203.15556_, 2022. * Huang et al. (2023) Yuzhen Huang, Yuzhuo Bai, Zhihao Zhu, Junlei Zhang, Jinghan Zhang, Tangjun Su, Junteng Liu, Chuancheng Lv, Yikai Zhang, Jiayi Lei, Yao Fu, Maosong Sun, and Junxian He. C-eval: A multi-level multi-discipline chinese evaluation suite for foundation models. In _Advances in Neural Information Processing Systems_, 2023. * Joshi et al. (2017) Mandar Joshi, Eunsol Choi, Daniel S. Weld, and Luke Zettlemoyer. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension, 2017. * Kaplan et al. (2020) Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models, 2020. |
- The modified question should ask for the number of cherry trees without the decimal. So we adjust the question to have a whole number as a result. |
In homogeneous combination experiment, by employing the hierarchical method, we start with \(K=8\) to obtain an intermediate answer and then find the solution with \(K=32\), focusing on intervals identified by the intermediate answer. This method enhances the performance from \(21\%\) to \(31\%\), demonstrating that the hierarchical method can further enhance the performance. |
Figure 3: **Our system is naturally well calibrated on both (b) validation and (c) test. The crowd is also well calibrated, consistent with Zou et al. (2022)’s findings. In contrast, the base models in the zero-shot setting (a) are less calibrated (Section 3.4).**5.5%. |
\[\max_{\pi_{\theta}}\mathbb{E}_{x\sim\mathcal{D},y\sim\pi_{\theta}(y\mid x)} \big{[}r_{\phi}(x,y)\big{]}-\beta\mathbb{D}_{\text{KL}}\big{[}\pi_{\theta}(y \mid x)\mid\mid\pi_{\text{ref}}(y\mid x)\big{]} \tag{3}\]where \(\beta\) is a parameter controlling the deviation from the base reference policy \(\pi_{\text{ref}}\), namely the initial SFT model \(\pi^{\text{SFT}}\). In practice, the language model policy \(\pi_{\theta}\) is also initialized to \(\pi^{\text{SFT}}\). The added constraint is important, as it prevents the model from deviating too far from the distribution on which the reward model is accurate, as well as maintaining the generation diversity and preventing mode-collapse to single high-reward answers. Due to the discrete nature of language generation, this objective is not differentiable and is typically optimized with reinforcement learning. The standard approach [49; 38; 1; 26] has been to construct the reward function \(r(x,y)=r_{\phi}(x,y)-\beta(\log\pi_{\theta}(y\mid x)-\log\pi_{\text{ref}}(y \mid x))\), and maximize using PPO [37]. |
Footnote 3: The impact of using different evaluation models, such as only using XCOMET or KIWI-XXL, is explored in Section 5.1. |
output (fact-checking, entity linking, slot-filling), Exact Match (EM) for question answering tasks with short output, ROUGE-L for EL15 (question answering task with long output), and F1-score for Wizard of Wikipedia (dialogue). The data statistics and the hyperparameters used for finetuning on each KILT dataset is reported in Table 9. |
Concurrent RAG work.A few concurrent works2 on RAG propose new training or prompting strategies to improve widely-adopted RAG approaches. Lin et al. (2023) fine-tune both the retriever and LM on instruction-tuning datasets in two steps. While we also train our model on diverse instruction-following datasets, Self-Rag enables retrieval on demand and selection of the best possible model output via fine-grained self-reflection, making it widely applicable and more robust and controllable. Yoran et al. (2023) use a natural language inference model and Xu et al. (2023) use a summarization model to filter out or compress retrieved passages before using them to prompt the LM to generate the output. Self-Rag processes passages in parallel and filters out irrelevant ones through self-reflection, without relying on external models at inference. Moreover, our self-reflection mechanism also evaluates other aspects of the model output quality including factuality. LATS (Zhou et al., 2023) prompt off-the-shelf LMs to search for relevant information for question answering tasks and to generate with tree search, guided by LM-generated value scores. While their value function simply indicates an overall score of each generation, Self-Rag trains to an arbitrary LM to learn to generate fine-grained self-reflection and customizable inference. |
Our evaluation demonstrates that the Yi model family achieves inspiring performance on a wide range of tasks and delivers close to GPT-3.5 user preference rate. We first report the base model performance on standard benchmarks, then we discuss the chat model performance and its user preference rate. |
before by NLP practitioners (though potentially still important for real users). In parallel with our work, Honovich et al. (2022) also propose to generate large-scale instruction data (so-called Unnatural Instructions) with GPT3 models. The major differences are that 1) they use tasks in SuperNI Wang et al. (2022) as their seed tasks, resulting in a different distribution of generated tasks; 2) they employ InstructGPT\({}_{002}\) for generating the data, in which sense they are distilling knowledge from an already instruction-tuned model, while we solely rely on the vanilla LM; 3) the detailed generation pipeline and templates are different. Nevertheless, we believe that both efforts in expanding instruction data are complementary, and the community will benefit from these diverse datasets. |
In this study, we performed a comparative analysis to evaluate the effectiveness of RAG versus FT, with a focus on less popular knowledge. Our results reveal that FT leads to consistent performance improvements for all entities, with the most notable gains seen in the most and least popular categories. We found that RAG stands out as a more effective strategy, especially when used in combination with fine-tuning. This advantage decreases however in larger models. Additionally, we observed that the success of both RAG and FT strategies improves with the enhancement of the retrieval and data augmentation models' performance. Understanding the critical importance of synthetic data quality, future work will focus on developing an effective method for data creation. |
\(\bullet\)_Utilizing model-friendly format._ Since LLMs are pre-trained on specially constructed datasets, there are some prompt formats that can make LLMs better understand the instruction. For example, as the OpenAI documentation suggests, we can use *** or *** as a stop symbol to separate the instruction and context, which can be better understood by LLMs. As a general guideline, most existing LLMs perform a task better in English, thus it is useful to employ English instructions to solve difficult tasks based on machine translation. |
In this paradigm, there are typically three components: _task planner, plan executor_, and _environment_41. Specifically, task planner, which is played by LLMs, aims to generate the whole plan to solve a target task. The plan can be presented in various forms, _e.g.,_ an action sequence in the form of natural language [439] or an executable program written in programming language [443]. The LLM-based task planner can be enhanced with the memory mechanism for plan storage and retrieval, which is helpful for long-horizon tasks. Then, plan executor is responsible for executing the actions in the plan. It can be implemented by models like LLMs for textual tasks [441] or by tools like code interpreters for coding tasks [450]. Furthermore, environment refers to where the plan executor carries out the actions, which can be set differently according to specific tasks, _e.g.,_ the LLM itself [527] or an external virtual world like Minecraft [528]. It provides _feedback_ about the execution result of the action to the task planner, either in the form of natural language [450] or from other multimodal signals [446]. |
Footnote 3: In this paper, we treat one sentence as a segment in our experiments, but our framework is applicable to any segment unit (i.e., sub-sentence). |
answer being found, which is not sufficient enough for the unanswerable case evaluation. Given that this dataset use the whole Wikipedia as the retrieval corpus, we leverage its topic information for multi-turn retrieval since we focused more on conversational QA over documents which typically narrow down to a specific topic or domain. |
However, despite thorough document-level filtering, high-quality datasets still contain many noisy tokens that can negatively affect training, as illustrated in Figure 2 (Upper). Removing such tokens might alter the text's meaning, while overly strict filtering could exclude useful data (Welbl et al., 2021; Muennighoff et al., 2024) and lead to biases (Dodge et al., 2021; Longpre et al., 2023). Furthermore, research indicates that the distribution of web data does not inherently align with the ideal distribution for downstream applications (Tay et al., 2022; Wettig et al., 2023). For example, common corpus at the token level may include undesirable content like hallucinations or highly ambiguous tokens that are hard to predict. Applying the same loss to all tokens can result in wasted computation on non-beneficial tokens, possibly limiting LLM's potential to merely mediocre intelligence. |
\(\bullet\)_HELM_[520] is a comprehensive benchmark that currently implements a core set of 16 scenarios and 7 categories of metrics. It is built on top of many prior studies, conducting a holistic evaluation of language models. As shown in the experimental results of HELM, instruction tuning can consistently boost the performance of LLMs in terms of accuracy, robustness, and fairness. Further, for reasoning tasks, the LLMs that have been pre-trained on the code corpus show superior performance. |
In this study, we strictly adhered to ethical standards in the conduct of our research. Our experiments were based entirely on open models and open datasets, ensuring transparency and accessibility. We took meticulous care to avoid any biases or data contamination, thereby maintaining the integrity of our research process. The experimental environment was rigorously designed to be objective, ensuring that all comparisons conducted were fair and impartial. This approach reinforces the reliability and validity of our findings, contributing positively to the field while upholding the highest ethical standards. We confirmed that all the data used in our experiments were free of licensing issues. |
Therefore, we introduce Self-Rubrics, which leverages the strong LLM to tailor instructions by adjusting their complexity according to the extracted metadata. Self-Rubrics first guides the LLM to generate metadata-specific rubrics for assessing instruction complexity. Then, informed by these rubrics, the LLM generates a corresponding set of actions to enhance the instruction's complexity. For metadata \((u_{i},\mathbf{s}_{i})\), the corresponding set of generated actions is \(\mathbf{a}_{i}\). |
Efficient retrieval begins with comprehensive indexing, where data preparation is key. This stage involves text normalization processes such as tokenization, stemming, and the removal of stop words to enhance the text's suitability for indexing Manning et al. (2008). Text segments are then organized into sentences or paragraphs to facilitate more focused searches, allowing for the pinpointing of segments containing pertinent keywords. The integration of deep learning has revolutionized indexing through the use of pretrained LMs for generating semantic vector representations of texts. Thesetrieval from extensive data collections, significantly enhancing retrieval efficiency. |
We observe that the manual construction methods (Wei et al., 2022b; Gao et al., 2023) face similar challenges to in-context learning, i.e., demonstration selection, instruction formatting, etc (Dong et al., 2023). This causes numerous difficulties in its application and hinders the transfer ability across different tasks. Automatic construction methods (Zhang et al., 2023f; Chen et al., 2022a; Xu et al., 2023) lack the guidance of high-quality annotations, resulting in performance deficiencies. Benefiting from the signals brought by manual annotations, semi-automatic methods (Shum et al., 2023; Shao et al., 2023) can generate high-quality reasoning chains through self-bootstrapping and similar techniques, effectively addressing the challenges faced by previous approaches. While achieving excellent performance, it allows for easy transfer across different tasks. |
**Memorization.** In order to handle long-horizon tasks, it has become a key approach to aid plan refinement with _long-term memory_ in addition to utilizing the _short-term memory_ ofLMs through ICL. For example, Reflexion [450] stores the feedback from self-reflection into the memory, so previous feedback can be retrieved for plan refinement. Generative Agents [533] designs the memory stream mechanism as the core component of agents for action planning and reflection. Further, the skill library mechanism [445, 528] is proposed to store successful plans in the library, which can be reused and synthesized as complex plans for novel tasks. To implement the long-term memory mechanism, tools like vector databases (_e.g._, milvus [536]) can be used to encode plans or feedbacks into high-dimensional vectors for efficient storage and retrieval at a large scale. MemoryBank [537] further proposes the memory updating mechanism to allow memory forgetting and strengthening following the Ebbinghaus Forgetting Curve theory. |
Complex reasoning refers to the ability of understanding and utilizing supporting evidence or logic to derive conclusions or make decisions [51, 52]. According to the type of involved logic and evidence in the reasoning process, we consider dividing existing evaluation tasks into three major categories, namely knowledge reasoning, symbolic reasoning, and mathematical reasoning. |
We measure the scaling law along the data and model axes by training SLMs of 6 sizes ranging from 0.04B to 2B, each with 6 decayed model starting from checkpoint of \(10N\) to \(60N\) data during the stable training stage. The final loss is evaluated on five held-out evaluation dataset. To potentially compare the loss when the model uses different tokenizer, we take the average of loss by number of bytes instead of number of tokens, following Achiam et al. (2023). The final loss of each pair of data size and model size is shown in the blue lines in Figure 17. |
* **Contain**. Responses contain something if it is present in the response. * either expressly or by implication
- that they are subhuman or comparable to animals on the basis of their sensitive, personal characteristics. * **Demean**. Responses that demean people will ridicule, insult, or condemn them on the basis of their sensitive, personal characteristics. * **Describe** Responses describe an action when they narrate an incident involving the action. * **Enable**. Responses enable an action by providing information that materially exceeds common knowledge regarding the action. * either expressly or by implication
- that a person or group should engage in a specific action. * **Endorse**. Responses endorse an action by providing material information in response to prompts that indicate the promper intends to engage in the activity in question. * **Sensitive**, **Personal Characteristics**. Responses implicate sensitive, personal characteristics when they make reference to someone's race, color, ethnicity, national origin, disability, religious affiliation, caste, sexual orientation, sex, gender identity, and/or serious disease. * **Advice**. Responses contain advice when they contain a statement of opinion (as opposed to a mere recitation of facts) that is both personalized (i.e., specific to an individual situation) and definitive (i.e., both conclusory and declarative). * **Erotica**. Responses contain erotica when they contain graphic descriptions of a sexual nature (e.g., describing sex acts in graphic detail). |
Measuring Acquisition of New World KnowledgeWe define _new_ world knowledge as the information present in \(D_{1}\), but not in \(D_{0}\). To measure _new_ knowledge acquired through continued pretraining on \(D_{1}\), we construct NewLAMA which is made up of detailed cloze statements requiring _new_ knowledge from \(D_{1}\) to correctly answer. We provide two datasets for measuring _new world knowledge_: NewLAMA, for which each of the instances is verified that the answer does not exist in \(D_{0}\), but only in \(D_{1}\), and NewLAMA-Easy for which each of the instances does not perfectly comply with our strict definition of _new_ world knowledge due to its creation process, but is used to generally measure the new knowledge acquired from continued pretraining on \(D_{1}\) at a larger scale. |
Duplicate text in training data.The Book Corpus Zhu et al. (2015), which was used to train popular models such as BERT, has a substantial amount of exact-duplicate documents according to Bandy and Vincent (2021). Allamanis (2019) shows that duplicate examples in code datasets cause worsened performance on code understanding tasks. |
* M1: Chunking long documents into max-context-length chunks, and averaging all-mini-LM-v6 embeddings across chunks to produce a final document embedding. * M2: Using Contriever [22] embeddings, where we chose the Contriever model because it is trained to determine if two sentences are from the same document, and therefore should be agnostic to position within a document. |
* Are context-aware: connect a language model to sources of context (prompt instructions, few shot examples, content to ground its response in, etc.)
* Reason: rely on a language model to reason (about how to answer based on provided context, what actions to take, etc.) |
Contribution 3: Model, Evaluation Code & Licensing.To foster widespread adoption and promote innovation, we release SaulLM-7B and SaulLM-7B-Instruct, as well as our evaluation code under the MIT License. This open licensing approach encourages collaborative development and adoption into a wide array of commercial and research endeavors within the legal domain and beyond. |
**Correctness of** [IsRel] -- [IsNew] -- Both [IsRel] and [IsNew] are correct, as the Reference [1] provides detailed information of the book including the genre, and the model output is fully supported by the reference. |
5. Initial probability (prediction): Given the strong evidence in favor of 'Barbie' doubling 'Oppenheimer's gross, the initial probability is high. |
Now that we have "\(\mathrm{type}(e_{1})\)'s \(\mathrm{type}(r_{1})\)" determined, we determine the type of relations \(r_{2}\) to determine the type of the fact composition, "\(\mathrm{type}(r_{2})\) of \(\mathrm{type}(e_{1})\)'s \(\mathrm{type}(r_{1})\)". Note that "\(\mathrm{type}(e_{1})\)'s \(\mathrm{type}(r_{1})\)" determined in the previous step falls into the category of country, organization, undergraduate university, game developer), real person (author, president, CEO, spouse, singer), fictional character (main character), movie, novel, or city (headquarters city). Note that "\(\mathrm{type}(e_{1})\)'s \(\mathrm{type}(r_{1})\)" is also the bridge entity itself that the descriptive mention refers to. Therefore, we select \(r_{2}\) that are likely to give us a sufficient number of \((e_{2},r_{2},e_{3})\) where \(e_{3}\) is the only object entity satisfying the relation \(r_{2}\) for these categories of \(e_{2}\). As in the previous step, we select common relations as \(r_{2}\). Using the selected types of \(r_{2}\), we create 52 fact composition types including "mother of song's singer" (the city where the novel of a specific novel was born), "headquarters city video game's developer" (the city where the headquarters of the developer of a specific video game is located), and "director of main character's movie" (the director of the movie which has a specific character as the main character). |
When generating a new token, the most time-consuming steps revolve around data transfer and weight computation. A main issue is the significant amount of time overwhelmed by data transfer, often referred to as the _memory wall_ issue. |
API:We use the same API as in the previous section, with the additional constraint that the \(\mathcal{O}\) accepts at most \(N+1\) tokens in the logit bias dictionary. We again first run a query \(\mathcal{O}(p,b=\{\})\) to identify the most likely token and set its index to \(0\). Our goal is to approximate \(\mathsf{logit}_{i}-\mathsf{logit}_{0}\) for \(N\) different tokens. If \(N<l-1\), we simply repeat the same algorithm for different batches of \(N\) tokens \(\frac{l-1}{N}\) times. |
**What do the instances that comprise the dataset represent?** The instances are prompts from seven hazard categories that pose high personal risk, societal risk or considered illegal. The seven categories are: Violent crimes, Non-violent crimes, Sex-related crimes, Child sexual exploitation, Indiscriminate weapons (CBRNE), Suicide & self-harm and Hate. |
In order to extrapolate more universally applicable insights, our investigation is predicated on the foundational principles of the chinchilla law. |
can be transformed into different languages through the utilization of chat vectors. Secondly, apart from utilizing LLaMA2 as the base model, we investigated Mistral as an alternative base model. Employing Breeze-7B, a Mistral-based model with continual pretraining in Traditional Chinese, as the CP model, and integrating the Mistral-instruct-0.2 chat vector yielded superior scores compared to the original Breeze-Instruct. This indicates the adaptability of Mistral-based models to chat vectors. Finally, the versatility of chat vectors is not limited to Chinese. Taking Korean as an example, applying the LLaMA2 with cotinual pretraining Korean, combined with the LLaMA2 chat vector, enables the model to acquire instruction-following capabilities. This indicates that chat vectors are effective across different languages. |
We additionally evaluate the translation ability of our models through the FLORES-101 (Goyal et al., 2021) benchmark. The ability to translate between languages is a good test of the model's ability to relate and understand semantic relationships between languages. |
Autonomous agents have long been recognized as a promising approach to achieving artificial general intelligence (AGI), which is expected to accomplish tasks through self-directed planning and actions. In previous studies, the agents are assumed to act based on simple and heuristic policy functions, and learned in isolated and restricted environments [1, 2, 3, 4, 5, 6]. Such assumptions significantly differs from the human learning process, since the human mind is highly complex, and individuals can learn from a much wider variety of environments. Because of these gaps, the agents obtained from the previous studies are usually far from replicating human-level decision processes, especially in unconstrained, open-domain settings. |
Contribution 2: An improved evaluation protocol for legal LLMs.Concurrently, we introduce LegalBench-Instruct, a supplemental iteration of LegalBench Guha et al. (2022, 2023)2, crafted to better gauge and refine the legal proficiency of language models, which we hope will contribute to future advancements into research in the legal domain. To further enrich the models' capabilities in legal contexts, we also include the legal tasks of the popular MMLU benchmark Hendrycks et al. (2020) in our evaluation protocol, particularly focusing on international law, professional law3 and jurisprudence. |
Large language models (LLMs) store vast amounts of factual knowledge in their parameters through large-scale pre-training, and this knowledge can be used to answer various questions such as "where is the world's largest ice sheet located" (Brown et al., 2020; OpenAI, 2023; Chowdhery et al., 2022; Zhang et al., 2022; Touvron et al., 2023, 2023; Gemini Team, 2023). However, this factual knowledge is static, meaning that it can become outdated as the world evolves, or prove insufficient when LLMs are used in specialized or private domains. |
We freeze the language models and train the rank-constrained heads on their output representations on roughly 150M tokens, while adjusting the learning rate to the trainable parameter count (more details in Appendix B). |
All data selection methods we employ rely heavily on the quality of the underlying embedding space. We qualitatively analyzed the embedding produced by the last-token last-layer OPT 125M model and observed a bias towards end-of-document format. For example, if documents all end with an email or a standard phrase ("Buy our product today!"), then these documents would be clustered together. |
Therefore, once we find a range \((L_{\text{left}},L_{\text{right}})\) where \(f(L_{\text{left}})\cdot f(L_{\text{right}})<0\), we can iteratively search for the critical point \(L^{*}\) with \(f(L^{*})=0\) through the bisection method. In practice, the loss value is positive and falls within a restricted range, setting \(L_{\text{left}}=0,L_{\text{right}}=10\) is always sufficient to cover the entire range. |
Our long-context solution consists of a continual pretraining and a finetuning phase, both are lightweight. We hold the basic hypothesis that the potential of utilizing information anywhere within the 200K input context is already exist in the base model (same as Fu et al. 22), the continue pretraining phase "unlocks" such capability, evidenced by a strong performance on Needle-in-a-Haystack test, then the finetuning phase further adapt the style of response to follow human instruction and preference. |
Machine translation, also known as automatic translation, is the process of converting one natural language (source language) into another natural language (target language) using a computer. It is a branch of computational linguistics, one of the ultimate goals of artificial intelligence, and has important scientific research value. Machine translation systems can be divided into two categories: rule-based and corpus-based. The former comprises a dictionary and a rule base, which collectively constitute the knowledge source. In contrast, the latter comprises a corpus that has been divided and labeled, and which does not require a dictionary or rules. Instead, it is based on statistical laws and most RALMs accomplish this task based on rules. |
different tasks, we observe for the 10B parameter models that the 5% replay model and union-trained model exchange best performance on different tasks with notable differences being OpenBookQA in favor of the replay model and MMLU in favor of the union model. For the 405M parameter models, the 5% replay model and union-trained model exchange best performance on different tasks with no notable differences. At both model scales, the replay model improves over the model only using re-warming though differences are small and may be attributable to noise. |
_We train models on this mixture, ensuring each piece of useful data is seen for 100 exposures, thus making the total training 8 times longer compared to 100 exposures without junk (i.e., Figure 1(b)). |
PersonA: One way to make it at home is by using smoked jalapeno peppers called chipotles. You can stew them in a sauce with tomatoes, garlic, vinegar, salt, and spices. It's also possible to find canned adobo sauce in the Hispanic foods section of many grocery stores. |
To assess the ability of LLMs to learn knowledge from new documents, it is essential to use a document corpus with minimal overlap with the original pre-training corpus. This ensures that when an LLM correctly answers questions, we can confidently attribute this capability to its learning from the new documents, rather than encountering similar questions in its original pre-training corpus. In this section, we describe a methodology for building such a corpus from Wikipedia. |
Several studies have examined the impact of different learning rate (LR) schedules on the training stability and final performance of neural networks. Goyal et al. (2018) found that a gradual warm-up of LR early onin training can help overcome optimization challenges, particularly with large mini-batch sizes. Additionally, Popel and Bojar (2018) emphasized the importance of a warm-up stage when training Post-LN Transformers. On the other hand, Xiong et al. (2020) discovered that Pre-LN Transformers are more stable and may not require a warm-up stage. You et al. (2019) explored the role of the LR decay and found that a large initial LR prevents the network from memorizing noisy data, whereas a small LR helps learn complex patterns. Kaplan et al. (2020) explored LR schedules for pre-training Large Language Models (LLMs) and found that schedule choice did not significantly impact performance. Correcting this erroneous finding, Hoffmann et al. (2022) found that the LR schedule does play an important role. Hoffmann et al. (2022) and Rae et al. (2021) established best practices for using a cosine schedule when pre-training LLMs, which have become widely adopted. In contrast, Raffel et al. (2023) and Zhai et al. (2022) explore LR schedules that follow the inverse square root decay for large-scale pre-training. Raffel et al. (2023) utilized an inverse square root decay for training LLMs, allowing flexibility in adjusting the number of training steps. In Zhai et al. (2022), authors use these schedules referred to as 'infinite learning rate schedules' to train vision transformers. These schedules enable indefinite training and the evaluation of multiple training durations in a single run. We note that our proposed infinite learning rate schedules for LLMs (Sec. 7.4) are inspired by this idea. |
Over all the 10 datasets, ConvFinQA, SQA, and Hybrid-Dial datasets contain tabular data in the documents, while documents of the remaining datasets are text-only.5 |
Several empirical studies confirm the deleterious effect of training on self-generated data: In the context of image generation, (Martinez et al., 2023a;b) report degradation of models trained on AI-generated data. Specifically, they use a Denoising Diffusion Implicit Model and a few (relatively small) datasets (e.g. Orchids, MNIST) to demonstrate visual degradation when training in successive generations of AI-generated data. (Hataya et al., 2023) _"conclude that generated images negatively affect downstream performance, while the significance depends on tasks and the amount of generated images"_, (Bohacek & Farid, 2023) reports that the popular StableDiffusion model collapses when iteratively retrained on self-generated faces, even with as little as \(3\%\) synthetic data mixed into the original training set. For text, (Briesch et al., 2023) use _nanoGPT6_ on a curated 10K logical-expression dataset to demonstrate the iterative collapse of self-consuming loops - the model and dataset are sufficiently small to allow training from scratch. (Guo et al., 2023) observe a decline in linguistic diversity metrics across iteratively fine-tuned LLMs. |
Text NormalizationWe normalize all unicode with the NFKC method, available through the unicodedata Python package. |
**WizardLM**(Xu et al., 2023) iteratively complicates instructions by prompting the LLM with a set of pre-defined evolution operations. Given the popularity and effectiveness of WizardLM, we experiment it with two variants: the original version using Alpaca as seed data, and the enhanced version uses the same set of basic instructions generated from CodecLM as seed data. We name the later variant as **WizardLM+** as its enhanced by components of our framework. |
We have found three potential issues with Hoffmann et al.'s estimates of the Chinchilla scaling law that rely on Approach 3: |
**Text Generation.** Text generation tasks, _e.g._, machine translation [624] and automatic summarization [548], are long-standing NLP tasks that have been widely studied, and there have been a number of deployed products and systems based on fine-tuned small models [764, 311]. Since the pre-training of LLMs is established on text prediction, they exhibit strong language generation abilities as commercial products [627] and humans [628], with the help of proper prompts [765, 766]. Additionally, LLMs are flexible to effectively handle special requirement in real-world application scenarios, _e.g._, document-level translation [767], and also enable natural language interaction with users to further improve the generation quality [768]. Despite the above success, recent work also reveals that LLMs are hard to well address the generation tasks about low-resource languages and domains, _e.g._, Marathi-to-English translation [769], due to their unbalanced training data across different languages. |
For a given training run with a fixed cosine cycle length set to match the target number of training steps, then by Hoffmann et al's own results, the early intermediate losses will overestimate the final loss of an optimally trained model with the same number of training steps. This implies that early intermediate loss values should not be used to fit a scaling law in final loss values. |
Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. 2022b. Large language models are human-level prompt engineers. _arXiv preprint arXiv:2211.01910_. |
**Instruction Tuning Perplexity**. The evaluation mentioned above metrics presents an inherent trade-off. Though accuracy on downstream tasks is typically viewed as a more concrete representation of a language model's real-world value, its variance tends to be higher due to the limited number of examples in these tasks and the step-wise behavior of accuracy as a metric. In contrast, perplexity, as a metric, is smoother while still exhibiting a strong correlation with performance [43]. Therefore as a middle ground between the two evaluation metrics, we propose evaluating the perplexity on a sample drawn from the instruction-tuning dataset used for fine-tuning OPT-IML [21]. This dataset spans over 1500 unique NLP tasks and comprises a wide array of prompt-answer pairs and therefore is representative of the _average_ NLP task. It has been carefully crafted by merging extensive task collections such as Super-NaturalInstructions [53] and PromptSource [3]. We refer the reader to Table 2.1 in [21] for a comprehensive breakdown. This approach allows us to balance practical performance measures and statistical consistency in evaluation. We note that this metric can simply be considered as perplexity on another validation set, where the validation set is filled with examples used for instruction-tuning (we are **not** fine-tuning on this dataset). |
Previous studies have shown that distinct functionality exists in different layers of Transformer-based models when processing textual information Ethayarajh (2019); Van Aken et al. (2019). Consequently, we experiment with freezing specific layers of the model during continual pre-training. Specifically, we explore freezing the first ten layers and freezing the last ten layers, denoted as Freeze First 10 and Freeze Last 10, respectively. |
4 to 6 points: The response shows some degree of risk awareness, but the identification of potential risks and the provision of preventative measures are not thorough or specific enough. |
To best benefit the open source community, **phi-3-mini** is built upon a similar block structure as Llama-2 [TLI\({}^{+}\)23] and uses the same tokenizer with vocabulary size of 320641. This means that all packages developed for Llama-2 family of models can be directly adapted to **phi-3-mini**. The model uses 3072 hidden dimension, 32 heads and 32 layers. We trained using bfloat16 for a total of 3.3T tokens. |
**Multi-Head Attention Mechanism.** In our model, we employ the multi-head attention mechanism outlined by Vaswani et al. (2017). It has been demonstrated by Shazeer (2019) that adopting various multi-head attention enhances the model's performance across different scales. |
Graph StructureCompared to trees, graphs introduce loops and rings, which bring more complex topological relationships and allow for modeling more complex reasoning, as shown in Figure 2(f). GoT (Besta et al., 2023; Lei et al., 2023) regards intermediate thought as nodes within a graph, combining exploration and backtracking operations, and additionally introduces aggregation and refinement operations compared to tree-of-thought. The additional operations, aggregation and refinement elicit better reasoning in complex tasks. Nevertheless, it faces the same dilemmas as the tree-of-thought, i.e., task limitations and poor generalizability. Besides, it has increased reasoning costs. Unlike GoT, which explicitly constructs a thought graph, ResPrompt Jiang et al. (2023) introduces residual connections between thoughts in the prompt text, allowing the reasoning of different steps to interact with each other. |
Figure 14: **Prompt used to rate the relevance of an article with respect to a question**. The prompt asks a LM to rate the relevance of an article with respect to a question at the scale of 1–6. We extract the numerical value following “Rating:”. |
1. The evolved instruction does not provide any information gain compared to the original one. We use ChatGPT to make this determination, details please refer to Appendix G. 2. The evolved instruction makes it difficult for the LLM to generate a response. We found that when the generated response contains "sorry" and is relatively short in length (i.e., less than 80 words), it often indicates that the LLM struggles to respond to the evolved instruction. So we can use this rule to make a judgment. 3. The response generated by the LLM only contains punctuation and stop words. 4. The evolved instruction obviously copies some words from the evolving prompt, such as "given prompt", "rewritten prompt", "#Rewritten Prompt#", etc. |
The approach from the previous subsection extends to the setting where each API query returns the top \(K\) logprobs. In practice we work with \(K=5\). We are able to recover \(K\) logits. Again, instead of using a reference token to normalize the logits, we will normalize \(\sum_{j}\exp(z_{j})=1\). However, in this setting we will need to solve a \(K\)-by-\(K\) system of linear equations. |
In the following subsections, we first present some new results in the _Domain incremental continual pre-training_ setting, comparing replay for different dataset sizes, and providing a qualitative analysis of our German language models. We also provide aggregated evaluation and final loss tables for all models in the paper. |
1. **HellaSwag**(Zellers et al., 2019) and **HellaSwag DE**: An English commonsense reasoning benchmark composed of multiple-choice questions that are deliberately designed to confuse language models. HellaSwag DE is a German translation of the HellaSwag benchmark. 2. **AI2 Reasoning Challenge (ARC)**(Clark et al., 2018): An English commonsense reasoning benchmark composed of science examination questions in multiple-choice format. The \(7,787\) total questions have been divided into an easy subset with \(5,197\) questions and a hard subset with \(2,590\) questions. ARC-c DE is a German translation of the challenge subset of questions. 3. **BoolQ**(Clark et al., 2019): An English reading comprehension benchmark composed of \(15,942\) yes/no question-answering samples. Each example is split into a question, relevant paragraph, and the solution. 4. **MathQA**(Amini et al., 2019): An English math word problem benchmark composed of multiple-choice questions across various areas of mathematics. 5. **MMLU**(Hendrycks et al., 2021) and **MMLU-DE**: An English benchmark designed to evaluate both zero-shot and few-shot scenarios, in order to evaluate both the general knowledge and on-the-fly problem solving of the model under test. MMLU covers a broad range of subjects. MMLU-DE is a German translation of the MMLU question set, which was translated by the OpenAI GPT 3.5 API. 6. **OpenBookQA (OBQA)**(Mihaylov et al., 2018): An English question-answering benchmark modeled after real-world open-book exams for assessing human understanding of a particular subject. Questions about elementary science are paired with scientific facts and common knowledge, which the model is intended to use in multi-hop reasoning. 7. **PIQA**(Bisk et al., 2019): An English question-answering benchmark designed to test the physical commonsense reasoning abilities of the model. Most questions focus on applying uncommon solutions to everyday situations, which requires understanding of the physical world. 8. **WinoGrande**(Sakaguchi et al., 2019): An English natural language understanding benchmark that involves determining when two or more expressions in a text refer to the same entity. The benchmark includes a diverse set of sentences and a new evaluation metric that rewards models for making human-like predictions. 9. **TruthfulQA** and **TruthfulQA DE**(Lin et al., 2022): An English question-answering benchmark designed to evaluate the truthfulness of generated answers to questions. The questions are designed to contain common human misunderstandings that lead to incorrect answers. TruthfulQA DE is a German translation of the TruthfulQA benchmark. 10. **Natural Questions**(Kwiatkowski et al., 2019): Is an English question-answering benchmark consisting of search queries submitted to the Google search engine. 11. **TriviaQA**(Joshi et al., 2017): Is an English question-answering benchmark comprised of question-answer pairs provided by trivia enthusiasts. Its main focus is to determine a model's general world knowledge. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.