Datasets:
Translation-EnKo
/
arxiv-translation

Dataset card Data Studio Files Community
2
text
stringlengths
4
4.47k
In this subsection, we compare the Mamba with Transformer, and CNN based backbone for the tracking task4 based on OSTrack [210]. Specifically, the CNN based trackers are TrDiMP [211], ToMP50 [212], DiMP50 [213], PrDiMP [214], KYS [215], and ATOM [216]; the Transformer based trackers are HDETrack [217], AiATrack [218], STARK [219], TransT [220], MiFormer [212], and SimTrack [222]. To achieve a fair comparison, we train and test these trackers on a large-scale event-based tracking dataset, EventVOT [217], which contains 841, 18, and 282 videos, respectively. The detailed experimental results are reported in Table XI and Fig. 13. Note that, three widely used evaluation metrics are used for the comparison, including Success Rate (SR), Precision Rate (PR), and Normalized Precision Rate (NPR). According to Table XI, we can find that the performance is slightly decreased when replacing the ViT using the Mamba backbone network, meanwhile, it brings about a huge reduction in the number of parameters (4.1M only). Therefore, we can draw the conclusion that the Mamba network will be a promising choice for the event-based tracking.
**Effectiveness of Core Designs.** We show component-wise contributions in our framework in Table 3. The 1st row has the result from Self-Instruct as a baseline; In the 2nd row, we only align the LLM with basic instructions from instruction metadata; We gradually add Self-Rubrics and Contrastive Filtering in the 3rd and 4th rows, respectively. We clearly observe that every component contributes to the final performance. Interesting, the performance of using basic instructions from metadata is even on par with that of WizardLM+ in Table 1. This observation indicates that human-crafted strategies for complicating instructions may not fit different types of instructions. On the contrary, Self-Rubrics adaptively generates instruction improving actions based on different metadata, resulting in better tailored instructions for the target LLM. Further improvements from Contrastive Filtering demonstrate that selected data are indeed more effective for alignment.
Chain-of-thought reasoning, a cognitive process fundamental to human intelligence, has garnered significant attention in the realm of artificial intelligence and natural language processing. However, there still remains a lack of a comprehensive survey for this arena. To this end, we take the first step and present a thorough survey of this research field carefully and widely. We use X-of-Thought to refer to Chain-of-Thought in a broad sense. In detail, we systematically organize the current research according to the taxonomies of methods, including XoT construction, XoT structure variants, and enhanced XoT. Additionally, we describe XoT with frontier applications, covering planning, tool use, and distillation. Furthermore, we address challenges and discuss some future directions, including faithfulness, multi-modal, and theory. We hope this survey serves as a valuable resource for researchers seeking to innovate within the domain of chain-of-thought reasoning1.
Our attack nearly perfectly extracts the model size of all models--except for GPT-2 Small where our extracted size of 757 is off by 11 from the correct 768. Why is this?
When the reversal curse was discovered in Allen-Zhu & Li (2023b), the authors utilized a biography dataset of 100K randomly generated individuals with unique English names. The biographies were either generated using a pool of sentence templates (the bioS dataset) or generated using the Llama model (the bioR dataset) (Touvron et al., 2023a). The biography entries always start with the person's full name.4 The reversal QA task is, therefore, very natural: given a person's partial or full biography details, ask for the person's name.
Lu et al. (2023) introduces multilingual dictionaries in prompts to enhance machine translation. Li et al. (2023) proposes chain-of-knowledge (CoK-Li), which obtains structured knowledge from a knowledge base via a query generator to perform knowledge-guided reasoning. Wang et al. (2023) (CoK-Wang) also retrieves structured knowledge from KB. Moreover, it estimates the reasoning chains in terms of factuality and faithfulness and prompts models to rethink unreliable reasonings, which mitigates the knowledge retrieval errors in CoK-Li. KD-CoT Wang et al. (2023) addresses factual reasoning problems through a multi-turn QA approach. They design a feedback-augmented retriever for retrieving relevant external knowledge in each round of QA to calibrate the reasoning process. Other studies use the model's own memory as external knowledge. For example, Memory-of-Thought Li and Qiu (2023) first performs pre-thinking to save the high-confidence thoughts into external memory, and during inference, it lets the LLM recall relevant memory to aid reasoning.
This exact-substring-matching criterion, while conceptually simple, is computationally prohibitive with naive (quadratic) all-pair matching. To improve the efficiency, we concatenate all the examples of the entire dataset \(D\) into a giant sequence \(\mathcal{S}\), and construct a Suffix Array \(\mathcal{A}\) of \(\mathcal{S}\). A suffix array [20] is a representation of a suffix tree [20] that can be constructed in linear time in \(\|\mathcal{S}\|\)[13] and enables efficient computation of many substring queries; in particular, they allow us to identify duplicated training examples in linear time. Suffix arrays have the advantage over suffix trees in that they are 10-100\(\times\) more memory efficient [21], requiring just 8 bytes per input token, though they are asymptotically less efficient for some query types. They have been used widely in NLP, such as for efficient TF-IDF computation [22] and document clustering [23].
This ensures that the loss is applied only to the tokens that are deemed most beneficial for the language model to learn from. In practice, token selection can be implemented by ranking the tokens in a batch according to their excess loss and using only the top \(k\%\) of tokens for training.
Retrieval-Augmented Generation (RAG) and Fine-Tuning (FT) stand out as two prominent approaches for adapting LLMs to specific domains. RAG retrieves relevant information from a document corpus and enhances LLM's response generation through the implementation of in-context learning (ICL). Conversely, the fine-tuning approach updates model weights to become adept at recalling specific information and enhance its memorization capabilities during inference. In the context of less popular knowledge, where limited data is available, data augmentation methods are utilized to generate synthetic training data, serving as an initial step towards fine tuning.
\(\bullet\)_Top-\(p\) sampling._ Since top-\(k\) sampling does not consider the overall possibility distribution, a constant value of \(k\) may be not be suitable for different contexts. Therefore, top-\(p\) sampling (_a.k.a._, nucleus sampling) is proposed by sampling from the smallest set having a cumulative probability above (or equal to) \(p\)[308]. In practice, the smallest set can be constructed by gradually adding tokens from the vocabulary sorted in descending order of generative probability, until their cumulative value exceeds \(p\).
Alpaca models show significant improvements over LLaMA.Among different settings, such as zero-shot or 5-shot, the Alpaca model series show significant improvements over LLaMA counterparts, demonstrating that the instruction-following models are more capable of handling these NLU-like tasks than pure language models. Unlike the phenomenon observed in the LLaMA series, we can see that Alpaca-Plus models yield significant improvement over basic Alpaca models. This might further indicate that instruction-following models are more capable of handling NLU-like tasks and can unlesah the power of using more pre-training data (LLaMA-Plus).
Since our paper focuses on the study of RefinedWeb alone, we chose to exclude common online sources of curated data from it. This serves two objectives: (1) it strengthens our results, by ensuring that RefinedWeb doesn't end-up actually being made mostly of known high-quality sources (e.g., Wikipedia represents a significant portion of C4); (2) future works may be interested in combining RefinedWeb with existing curated copora, which would require further deduplication if they are included in RefinedWeb. Accordingly, we remove common sources used in The Pile (Gao et al., 2020) from RefinedWeb. The full list of curated data sources domains that we blocked is in Table 14.
BART [55] uses a standard sequence-to-sequence translation model architecture. It is pre-trained by corrupting text with an arbitrary noising function, and then learning to reconstruct the original text.
Figure A.16: **Prompt for the auto-evaluation of qualitative criteria.** We constructed few-shot prompt with self-generated explanations alongside their respective dialogues as exemplars to evaluate and rate a new dialogue. This few-shot prompt included one example for each point on the 5-point rating scale.
2. Sprint Intervals: Sprint intervals involve short, all-out sprints followed by brief jogging or rest periods. This workout can improve speed and power.
**Setting:** Recall that there is an unknown vector \(z=\mathbf{W}\cdot g_{\theta}(p)\in\mathds{R}^{\ell}\) (i.e., the logits for a given prompt \(p\)) that we want to recover. We can make multiple queries to the API with the same prompt \(\mathcal{O}(p,b)\). Each query is specified by a vector \(b\in\mathds{R}^{\ell}\) (a.k.a. the logit bias).
We acknowledge the limitations of CodecLM from the following aspects to inspire future research opportunities in the field of LLM alignment.
In Table 5, we find that using "top-5" chunks as the context for training leads to improvements on the datasets that need retrieval. But it downgrades the performance on non-retrieval datasets. Overall, these two models perform comparable. It is because incorporating "top-5" retrieved chunks in stage-2 tuning aligns with the inference stage where retrieval is needed and hence, improves the Avg-ret score. However, mixing continuous and discontinuous documents could make the stage-2 tuning less stable, leading to sub-optimal results on non-retrieval datasets. We believe more future work can be conducted regarding the balance of incorporating continuous context and top-\(k\) retrieved chunks in the stage-2 tuning.
We also conduct experiments by freezing specific modules of the model during continual pre-training. We aim to explore whether these designated modules preserve the abilities acquired during the alignment operations.
Figure 1: Correlation between subject entity popularity in a question and the effects of RAG and FT on FlanT5-small performance in open-domain question answering. FT markedly improves accuracy in the initial and final buckets relative to others (indicated by the pink line).
1. Divine Rivals is set in a world affected by a war between two divine families, the Skywards and the Underlings. 2. The protagonist of "Divine Rivals" is Iris Winnow, an orphan who works as a junior reporter at the Oath Gazette in the city of Oath. 3. Iris Winnow's aspiration to become a renowned journalist is driven by her desire to uncover the truth about the war that resulted in the death of her mother and brother, Forest. 4.The conflict between the Skywards and the Underlings has caused chaos in the realm of Cambria, a place where magic and mythology interact with the mortal world. 5. Iris Winnow has a professional rival named Roman Kitt, a skilled storyteller from a wealthy family. 6. Roman Kitt is under pressure from his father to join the family business. 7. Zeb Autry, the editor of the Oath Gazette, does not want to cover the war, against Iris Winnow's wishes. 8. After the loss of her mother and brother, Iris Winnow's family is facing financial and personal struggles. 9. The Inkridation Tribune, a competing newspaper, advertises a job for war correspondents, leading both Iris Winnow and Roman Kit to compete for the position. 10. Iris Winnow maintains a mysterious correspondence with an unidentified individual who leaves her letters in her wardebro. 11. Iris Winnow's brother, Forest, introduced her to journalism. 12. Iris Winnow and her brother Forest had differing views on the war and the Skyward goddess, Enva. 13. Despite various challenges, Iris Winnow achieves her goal of becoming a war correspondent for the Inkridden Tribune. 14. Iris Winnow embarks on a dangerous journey to cover the war. 15. Iris Winnow continues to write letters to the unknown person who becomes her confidante during her journey. 16. The town of Avalon Bluff is preparing for a battle against an unknown enemy. 17. Dacre, the god of the Underlings. 18. Amidset the chaos, Iris Winnow and Roman Kitt rekindle their connection, and Roman Kitt proposes to Iris Winnow. 19. Iris Winnow experiences the horrors of war firsthand, including the loss of innocent lives and the destruction of her childhood home, Clover Hill. 20. Despite the hardships, Iris Winnow continues to report on the war, using her writing to document the battles and search for her brother, Forest. 21. Divine Rivals explores themes of identity, courage, and the power of words during war. 22. Iris Winnow's determination to make a difference by writing about the truth and seeking answers about her brother is a central theme in "Divine Rivals". 23. The author of "Divine Rivals" acknowledges her literary agent, film agent, subrights team, editors, cover designer, copy editor, and fellow authors for their assistance in publishing the book.
* LoraHu et al. (2022): A method that incorporates trainable low-rank decomposition matrices into each layer of the Transformer-based model. In our implementation, we selectively adapt only the query and value projection matrices of each layer in the model. * (IA)\({}^{3}\)Liu et al. (2022): A technique involving element-wise multiplication of the model's activations with a learned vector. We rescale the key and value matrices in self-attention modules and the inner activations in feed-forward modules in each model layer.
Additionally, our findings regarding empathic communication could also be partially attributed to the fact that AMIE responses were significantly longer than clinician responses (shown in Figure A.11), and presented with greater structure. This could potentially suggest to an observer that more time was spent preparing the response, analogous to known findings that patient satisfaction increases with time spend with their physicians [79, 80, 81].
In this section, we examine two widely used frameworks for knowledge injection: fine-tuning (FT) and retrieval augmented generation (RAG). We begin by formulating the knowledge injection problem, aiming to explain both methods using consistent terminology.
**Stage 2 (new output embeddings):** Our goal is to enhance the model's proficiency in accurately generating new tokens across various contexts by solely adjusting the output embeddings (lm_head). The decision to freeze all other parameters stems from the model's current unstable state. Allowing both input and output embeddings to be trained simultaneously would complicate achieving convergence, thus hindering the model's progress toward optimal performance. By freezing most of the parameters, we achieve more stable convergence. Moreover, this approach significantly reduces the training time, as it eliminates the necessity for back-propagation through the other layers.
AI Agent Agent Architecture AI Reasoning Planning Tool Calling Single Agent Multi Agent Agent Survey LLM Agent Autonomous Agent
Large Language Models.Recently, LLMs have demonstrated remarkable achievements across a broad spectrum of tasks. Prominent tech companies have made significant strides in developing highly proficient LLMs. These include OpenAI's GPT3&4 [1; 2], Google's PaLM [3; 4], and Bard3, DeepMind's Chinchilla [5], and Gopher [6], as well as Anthropic's Claude4. However, it is important to note that these models are closed-source and can only be accessed through specific APIs or may not be accessible at all.
This experiment may underestimate GPT-4's ability. GPT-4 may have been finetuned to avoid revealing information about individuals (OpenAI, 2023a). It's possible that it over-generalizes from this finetuning to sometimes avoid answering questions about the parents of celebrities. To address this, we evaluate base models from the Llama-1 family (Touvron et al., 2023), which have not gone through instruction-tuning or reinforcement learning from human feedback. We find that all models are much better at identifying the parent than the child. See Figure 5. Further details for Experiment 2 are in Appendix C.
Current Events ResultsThe evaluation on the current events task is shown in Table 2. RAG proves particularly effective due to the one-to-one correspondence between the questions and the auxiliary dataset (see Section 4.3). Fine-tuning is not competitive with RAG. However, fine-tuning with multiple paraphrases still provides a significant improvement over the baseline. We note that combining RAG with fine-tuning shows inferior performance compared to RAG alone.
Firstly, it can be observed that applying a monolingual tokenizer to multilingual data results in significantly higher fertility and parity scores (see Fig. 0(a) and Fig. 2). While multilingual tokenizers have lower fertility than monolingual English tokenizers on all non-English documents by a large margin, they are only slightly worse on tokenizing English documents, as shown in Fig. 0(b).
Using simulated dialogues allowed us to address the limited availability of high-quality, labelled real-world conversation data and improved the model's generalization and adaptability to diverse medical contexts. By leveraging this self-play paradigm, AMIE could continuously learn and refine its conversational and diagnostic capabilities during patient interactions.
Note that the above may not be feasible due to the absence of such aligned models, or the fact that it renounces control over the reference model, leading to safety concerns. Instead, we propose'step-wise DPO', named sDPO, where we use the preference datasets (or subsets of a preference dataset) in _a step-by-step manner_ when undergoing DPO training. The aligned model in the previous step is used as the reference model for the current step, which results in utilizing a more aligned reference model (_i.e._, a better lower bound).
Exact match (EM) is another metric that is mostly concerned with exact matches from (pre-defined) answers. It counts a prediction as correct if it exactly matches one of more than one desired reference text token by token. In some cases, it can be the same as accuracy and the equation 5 shows the mathematical definition. Here M is total number of correct answers and N is the total number of questions [202].
As illustrated in Figure 15, the generation process of CoT follows a chain structure in the basic CoT prompting approach, where LLMs generate CoTs step by step. Typically, CoT takes the format of natural language text. However, textual CoTs may not work well on complex tasks that require rigorous logic for reasoning. Considering this, some work uses code [506, 507] due to its structured and precise nature. Furthermore, the authors in [508] propose to dynamically select text or code as the format of CoTs to combine their advantages.
\(\bullet\)**Multi-modal learning using SSMs architecture.** Early multi-modal related works focused on how to learn modality-specific and modality-shared representations. Influenced by the Transformer network, current multi-modal algorithms usually directly encode and fuse the multiple cues in a unified Transformer network [267, 268]. Thus, the cost of the inference phase may be twice compared with a single modality only. How to design new SSMs-based backbones for cost-sensitive multi-modal learning is an important research topic.
Figure 12 \(|\) R-GSM example where the original problem can be correctly solved by GPT-4 Turbo, but the model fails on the reordered one.
One conversational QA sample is based on a document that we provide.
As observed in Table 2(b), all models perform poorly in the baseline setting. We intuit that models require detailed contexts and up-to-date information to make accurate forecasts. Our system addresses this issue via news retrieval and elicits better reasoning via optimized prompting strategies and fine-tuning.
methodology, and aims to be open to audit. For our core target audiences we anticipate the following specific requirements for the grading system: For **model providers** the benchmark should incentivize targeted safety improvements to models. Scores should be fair, meaning that they should not favor or disadvantage models of specific providers. Moreover, scores that indicate good safety performance should be achievable. For **model integrators** the benchmark should have a signal. A better score should mean a safer model. It should be possible to break down scores into more granular insights and differences between scores should be clear. For **AI standards makers and regulators** the benchmark should encourage wide participation and drive improvements in safety. It should also facilitate meaningful dialogue within the AI research and developer community.
Our proposed approach is to learn a conditional language model that maps from next-token probabilities back to tokens: \(p(x_{1:T}\mid\mathbf{v})\). We parameterize this distribution using a pretrained Transformer language model and train on samples from the unconditional model. Following work from Dumoulin et al. (2018) on feature-level conditioning, we use the cross-attention in an encoder-decoder Transformer to condition on the next-token vector.
Not that there's a thing wrong with frozen burgers. The key here is the meat seasonings, which are pretty strong and spicy and just GOOD, something else I think is really necessary in a turkey burger because ground turkey otherwise can be kind of flavorless. You'll need ground turkey, onion powder, chili powder, salt, pepper, and cyeyneme pepper for the burgers. Then the mayo takes garlic and onion. Then we need buns, clearly, swiss cheese, lettuce, and onion. I LOVE tomatoes but sometimes find that they get in the way of other flavors, so I left them off of this burger. Add them if you'd like to your array of toppings! First, we'll make the mayo. Grate the garlic directly into the mayo, add a pinch of salt, and squeeze in the lemon juice. Stir. Done! I love this. Then, we'll work on the burgers. Preheat a large skillet to medium-high heat with some olive oil, preheat the broiler to high, then add all the spices to the ground turkey.
To evaluate the learned model, we compute the loss on validation data \(\mathcal{D}_{\textbf{val}}\).
Memorizing training data.The privacy risks of data memorization, for example the ability to extract sensitive data such as valid phone numbers and IRC usernames, are highlighted by Carlini et al. (2020). While their paper finds 604 samples that GPT-2 emitted from its training set, we show that _over_\(1\%\) of the data most models emit is memorized training data. In computer vision, memorization of training data has been studied from various angles for both discriminative and generative models (e.g. Arpit et al., 2017; Webster et al., 2019; Feldman and Zhang, 2020; Stephenson et al., 2021; Tetewak et al., 2021).
Yes, we applied extensive preprocessing and cleaning of the data. We first filter URLs to remove adult content using a blocklist and a score system (Appendix G.1), we then use trafilatura (Barbaresi, 2021) to extract content from pages, and perform language identification with the fastText classifier from CCNet (Wenzek et al., 2020). After this first preprocessing stage, we filter data using heuristics from MassiveWeb (Rae et al., 2021) and our own line-wise corrections (Appendix G.2). Finally, we run extensive deduplication, removing URLs revisited across dumps (Section 3.3) and performing subsequently fuzzy and exact substring deduplication, with each stage drawing from Lee et al. (2022). See Section 3 for further details and Table 2 for an outline.
The ViT-S [19] and Mamba-based network Vim-S [61] are adopted as the backbone for this experiment. We follow the vision-language fusion based PAR framework VTB [207] which takes the pedestrian image and attribute set as the input and predicts the logistic scores of each attribute. From the experimental results reported in Table X, we can find that the Vim-S based PAR model achieves 81.08/73.75/80.91/84.96/82.52 on the PETA dataset, and 80.41/78.03/85.39/88.37/86.39 on the PA100K dataset. These results are significantly better than the ViT-S based model, but still significantly inferior to the compared PAR algorithms developed based on the Transformer network. For example, the ViT-B based VTB achieves 85.31/79.60/86.76/87.17/86.71, 83.72/80.89/87.88/89.30/88.21 on the PETA and PA100K datasets.
Each query may receive multiple answers (namely, the \(K\) largest \(a_{i}(z,b)\) values). For notational simplicity, we denote multiple answers to one query the same way as multiple queries each returning one answer. Suppose queries \(b^{1},\cdots,b^{m}\) were asked and we received \(m\) answers \((i_{1},a_{i_{1}}(z,b^{1}))\leftarrow\mathcal{O}(p,b^{1}),\cdots,(i_{m},a_{i_{m }}(z,b^{m}))\leftarrow\mathcal{O}(p,b^{m})\).
Footnote 14: _This summary includes a description of who the author thanks at the end of the book which is not important to the plot of the book.”_
For a given source sentence \(x\), whether translated from or to English, we utilize both GPT-4 and ALMA-13B-LoRA to generate respective translations, denoted as \(y_{\text{grt-4}}\) and \(y_{\text{alma}}\). Together with the original target reference \(y_{\text{ref}}\), this forms a triplet \(\mathbf{y}=(y_{\text{ref}},y_{\text{grt-4}},y_{\text{alma}})\), representing three different translation outputs for the input \(x\). The reference-free evaluation models KIWI-XXL and XCOMET are then employed to score these translations, with the average scores represented as \(\mathbf{s}=(s_{\text{ref}},s_{\text{grt-4}},s_{\text{alma}})\).3 The highest-scoring translation is labeled as the preferred translation \(y_{w}\), and the lowest-scoring as the dis-preferred translation \(y_{l}\), i.e., \(y_{w}=\mathbf{y}_{\arg\max_{i}(\mathbf{s})},y_{l}=\mathbf{y}_{\arg\min_{i}( \mathbf{s})}\), where \(i\) represents the index in the triplet. Translations with intermediate scores are not considered. An illustrative example of this selection process is depicted in Figure 3. It is important to note that even the dis-preferred translations may be of high-quality. The designation 'dis-preferred' indicates that there is still room for improvement, perhaps through the addition of minor details. This approach of using high-quality but not flawless translations as dis-preferred data aids in training the model to refine details and achieve perfection in generated translations.
\(\bullet\)**PKU-SafeRLHF**[181] encompasses more than 330K instances of expert comparison data, concentrating on the helpfulness and harmlessness. Each instance in the dataset includes a question and two responses, accompanied by safety labels for each response and two preference annotations between the two responses according to helpfulness and harmlessness. The harmlessness of a response indicates its classification as risk-neutral across all 14 harm categories, while the helpfulness of a response is evaluated based on its effectiveness in addressing the question.
In practice, implementing the suggested architecture is quite straightforward: Given an auxiliary knowledge base \(\mathcal{B}_{\mathcal{Q}}\) and a pre-trained embedding model \(\mathcal{M}_{e}\), we create a dense vector representation (embedding) per document \(b\in\mathcal{B}_{\mathcal{Q}}\) and store these in a vector store. Upon receiving a new query \(q\), we use its embedding, \(\mathcal{M}_{e}(q)\), to retrieve \(q\)'s top-\(K\) closest neighbors, \(\mathbf{b}_{q}=\{b_{k}\}_{1}^{K}\), according to dot-product ranking. We then update \(q\) to be \(\tilde{q}=\mathbf{b}_{q}\|q\), where \(\|\) denotes string concatenation. Finally, we return \(\mathcal{M}(\tilde{q})\) as the model's output.
**LM-guided CoT Performance.** Table 2 shows a comprehensive breakdown of our method's performance. Additionally, we explore an extension of our approach, which involves sampling multiple reasoning paths and subsequently ranking the most optimal rationales. Our method with only KD outperforms the original CoT prompting with 2% gain for HotpotQA and 10% for 2WikiMultiHopQA, respectively. Figure 2 illustrates the respective rationale qualities of all prompting techniques, reinforcing the effectiveness of our method in enhancing both answer prediction and rationale qualities. When employing the original CoT prompting for questions with lengthy contexts, models frequently recycle sentences from the provided context and struggle to deliver a conclusive answer to the question. This trend is mitigated by our approach, resulting in a significant decrease in error rates. It also surpasses the performance of CoT prompting + SC and is on par with standard prompting in terms of EM and F1. For the answer inclusion score, LM-guided CoT prompting is slightly higher (1-2%) than standard prompting. Furthermore, LM-guided CoT prompting + SC achieves the highest performance across all settings.
1. Unstable Anjnana 2. Acute Myocardial Infarction 3. Stable Anjnana 4. Gastroesophageal Reflux Disease (GERD) 5.
1. For LLaMA/Mistral models on \(\mathsf{bio5}(N)\) data, aiming to present _negative_ results, we select the best learning rate from three options in each data setting: * For \(N=50K\), we use \(wd=0.01\), \(lr=0.0003/0.0005/0.001\), and batch size 12 with bf16; * For \(N=100K\), we use \(wd=0.01\), \(lr=0.0003/0.0005/0.001\), and batch size 24 with bf16; * For \(N=200K\), we use \(wd=0.01\), \(lr=0.0003/0.0005/0.001\), and batch size 48 with bf16; * For \(N=500K\), we use \(wd=0.01\), \(lr=0.0002/0.0003/0.0005\), and batch size 96 with bf16; * For \(N=1M\), we use \(wd=0.01\), \(lr=0.0002/0.0003/0.0005\), and batch size 192 with bf16; * For \(N=2M\), we use \(wd=0.01\), \(lr=0.0003/0.0005/0.001\), and batch size 384 with bf16; * For \(N=5M\), we use \(wd=0.01\), \(lr=0.0003/0.0005/0.001\), and batch size 768 with bf16; * For \(N=10M\), we use \(wd=0.01\), \(lr=0.0003/0.0005/0.001\), and batch size 1536 with bf16; * For \(N=20M\), we use \(wd=0.002\), \(lr=0.0003/0.0005/0.001\), and batch size 1536 with bf16. (For \(N\leq 1M\), we also tested the same settings with fp16, finding similar results. However, LLaMA/Mistral models tend to fail more often with fp16, so we primarily used bf16.) 2. For GPT2\({}_{1/4}\): * For \(N=50K\), we use \(wd=0.01\), \(lr=0.0005/0.001\), and batch size 12 with fp16; * For \(N=100K\), we use \(wd=0.01\), \(lr=0.0005/0.001\), and batch size 24 with fp16; * For \(N=200K\), we use \(wd=0.01\), \(lr=0.0005/0.001\), and batch size 48 with fp16; * For \(N=500K\), we use \(wd=0.01\), \(lr=0.0003/0.0005\), and batch size 96 with fp16; * For \(N=1M\), we use \(wd=0.01\), \(lr=0.0003/0.0005\), and batch size 192 with fp16. 3. For GPT2\({}_{0}\), to present a _negative_ result, we use the same settings as in Parameter 4(a): * For \(N=50K\), we use \(wd=0.01\), \(lr=0.0003/0.0005/0.001\), and batch size 12 with bf16; * For \(N=100K\), we use \(wd=0.01\), \(lr=0.0003/0.0005/0.001\), and batch size 24 with bf16; * For \(N=200K\), we use \(wd=0.01\), \(lr=0.0003/0.0005/0.001\), and batch size 48 with bf16; * For \(N=500K\), we use \(wd=0.01\), \(lr=0.0002/0.0003/0.0005\), and batch size 96 with bf16; * For \(N=1M\), we use \(wd=0.01\), \(lr=0.0002/0.0003/0.0005\), and batch size 192 with bf16.
**Assumptions.** In Lemma 4.1, we assume that both the matrix with columns \(g_{\theta}\left(p_{i}\right)\) and the matrix \(\mathbf{W}\) have rank \(h\). These matrices have either \(h\) rows or \(h\) columns, so both have rank at most \(h\). Moreover, it is very unlikely that they have rank \(<h\): this would require the distribution of \(g_{\theta}\left(p\right)\) to be fully supported on a subspace of dimension \(<h\) across all \(p_{i}\) we query, or all \(h\ll l\) columns of \(\mathbf{W}\) to lie in the same \((h-1)\) dimensional subspace of \(\mathbf{R}^{l}\) (the output space of logits). In practice we find this assumption holds for all larger models (Table 2) and when different normalization layers are used (Appendix B.1).
Fig. 10: The probability distribution over the vocabulary in descending order for the next token of the context “_I am sleepy. I start a pot of”_. For ease of discussion, this example is given in word units instead of subword units.
The above is the formulation of our CPO loss, which includes one preference learning term \(\mathcal{L}_{\text{prefer}}\) and one negative log likelihood term \(\mathcal{L}_{\text{NLL}}\).
Below is a conversation with a helpful and terse assistant. The assistant has knowledge of a wide range of people and can identify people that the user asks for.
Figure 3: Examples from three datasets demonstrating differential LLM responses across various types of context modifications. Responses in red indicate wrong responses (different than the answer); responses in green indicate correct responses.
\(\bullet\)_Prompting Engineering_. Due to the strong language comprehension capabilities, people can directly interact with LLMs using natural languages. This introduces a novel strategy for enhancing agent capabilities, that is, one can describe the desired capability using natural language and then use it as prompts to influence LLM actions.
We can apply this reparameterization to the ground-truth reward \(r^{*}\) and corresponding optimal model \(\pi^{*}\). Fortunately, the Bradley-Terry model depends only on the difference of rewards between two completions, i.e., \(p^{*}(y_{1}\succ y_{2}\mid x)=\sigma(r^{*}(x,y_{1})-r^{*}(x,y_{2}))\). Substituting the reparameterization in Eq. 5 for \(r^{*}(x,y)\) into the preference model Eq. 1, the partition function cancels, and we can express the human preference probability in terms of only the optimal policy \(\pi^{*}\) and reference policy \(\pi_{\text{ref}}\).
Unlike fields such as science and medicine, the legal landscape varies significantly across countries and jurisdictions, reflecting differences not only in local laws but also in legal traditions, like common law versus civil law Henderson et al. (2022). Thus, we gathered legal texts from various jurisdictions, with a primary focus on the English language due to its widespread use in legal contexts worldwide. Our collection includes data from the U.S. Tuggener et al. (2020), Europe Chalkidis et al. (2019), and Australia Butler (2023), covering a diverse range of legal systems. Through this thorough curation process and aggressive cleaning (see Section 3.1.2), we end up with a corpus of 30 billion tokens, capturing the intricacies of legal language across regions.
struction data generated by real human users, OpenAI's LLMs (e.g., InstructGPT [2] and ChatGPT 4) have achieved great success. These open-domain instructions can fully unleash the unlimited potential of LLMs [14, 15, 16, 17] and enable them to perform more complex and diverse tasks. However, using humans to create open-domain instruction datasets like OpenAI did will encounter the following challenges. The whole annotating process is extremely expensive and time-consuming [18, 19, 20, 21]. On the other hand, the difficulty level distribution of human-created instructions is skewed towards being easy or moderate, with fewer difficult ones (according to the difficulty statistics of ShareGPT [22] from Figure (a)a). Possible reasons for this are that the proportion of experts among annotators is low and creating complex instructions demands a lot of mental effort. Human annotators are prone to fatigue and cannot sustain high-intensity work to produce a sufficient proportion of high-difficulty instructions [23, 24, 25, 26]. Based on these issues, developing an automatic method that can mass-produce open-domain instructions (especially the more difficult ones) at a relatively low cost becomes the key to further advancing instruction-tuned language models [27, 28, 29, 30].
**Is it possible to identify the individuals directly or indirectly from the dataset?** No. No individuals are mentioned.
7. Expressing support for people who intentionally engage in activities that cause harm. 8. Spreading information about how to engage in activities that cause harm. 9. Creating resources or assets that can be used in activities that cause harm.
**Quantization Libraries**.
Legal SourcesWe combine both previously available datasets, such as the FreeLaw subset from The Pile Gao et al. (2020) and MultiLegal Pile Niklaus et al. (2023), as well as data scraped from publicly available sources on the Web. We list the different sources of data in Table 1.
In this work, we aim to compare two choices for \(\mathcal{F}\): fine-tuning and RAG to see which option performs better in this problem.
Challenges in methods and data accessThe risks of harm created by AI systems are often difficult to identify, and their likelihood and severity cannot be easily estimated without extensive access to production systems and considerable resources [138, 139, 140]. Adoption of generative AI tools has been rapid but recent and, in part due to the novelty of these systems, we are unaware of longitudinal, quantitative and representative studies on how AI interactions lead to harm as of this writing. However, there is a growing body of evidence relating to individual incidents of harm that are associated with AI systems. Examples include giving potentially harmful diet advice to people at risk of eating disorders;26 inventing non-existing case law when asked to help draft legal briefs;27 and causing financial harm through overcharging customers.28 Some organizations have also released data from 'the wild' that provide insight into hazards created by real-world interactions with models [69, 70, 141]. However, accessing such data can be difficult for safety research given its sensitivity and the fact that it is mostly held by private companies.
Despite the performance, for training efficiency, our model is trained on 7K data for 10 epochs, while RFT-100 require training on 46K data for 3 epochs to achieve similar results. Notably, our approach achieves comparable performances without requiring extra machine computing resources.
data sources and model abilities. To enhance specific skills such as mathematics and coding in LLMs, or to develop specialized LLMs, a practical way is to employ a multi-stage training approach, _e.g.,_ general and skill-specific data can be scheduled at two consecutive stages. This approach of training LLMs on varying sources or proportions of data across multiple stages is also known as "data curriculum", which will be introduced below.
ination in Llama 27 and because the Tulu mix was primarily designed for Llama models - we will investigate the cause of this gap in future work. Overall, we see that OLMo greatly benefits from additional tuning and serves as a strong base model for downstream applications.
Results.We make the following observations from the results in Table 3. Self-Instruct boosts the instruction-following ability of GPT3 by a large margin. The vanilla GPT3 model basically cannot follow human instructions at all. Upon manual analysis, we find that it usually generates irrelevant and repetitive text, and does not know when to stop generation. Compared with other models that are not specifically trained for SuperNI, GPT3SEL-INST achieves better performance than T0 or the GPT3 finetuned on the T0 training set, which takes tremendous human labeling efforts. Notably, GPT3SEL-INST also nearly matches the performance of InstructGPT001, which is trained with private user data and human-annotated labels.
Metadata-conditioned instructions lay the groundwork for aligning the target LLM to desired tasks. Studies suggest that more complex instructions can improve alignment performance Xu et al. (2023); Zhao et al. (2023). A common practice is to involve human experts crafting general guidance to complicate instructions, such as adding reasoning steps or constraints. However, this one-size-fits-all strategy falls short for diverse instructions. Tailoring guidance to different tasks, like solving calculus problems versus writing news articles, requires distinct approaches.
We now evaluate our pre-trained language models on various zero-shot question answering (QA) benchmarks using the LLM Evaluation Harness1Gao et al. (2023).
We perform extensive experiments with an encoder-decoder model, T5 (Raffel et al., 2019), a large LM (\(\sim\) 737M params) initially pretrained on April 2019 dump of C4 and May 2020 dump of Wikipedia (thus \(D_{0}\) in our experiments) with salient span masking (SSM). The details of the pretraining, continual pretraining, and evaluation configurations are in Appendix C. We establish the following methods as the baselines for the CKL benchmark and categorize them into _regularization_, _rehearsal_, and _parameter-expansion_ methods. The specific hyperparamters used for the implementation of each method are detailed in Appendix D.
\(\bullet\)_P3_[182] is composed of 170 English NLP datasets and 2,052 English prompt templates, where the input and output of each data example have been formatted with specific prompt templates for composing the training instance.
The performance-vs-loss curves for all the intermediate checkpoints are shown in Figure 5. The trend is the same as Figure 2, but with larger variance.
At the heart of medicine lies the physician-patient dialogue, where skillful history-taking paves the way for accurate diagnosis, effective management, and enduring trust. Artificial Intelligence (AI) systems capable of diagnostic dialogue could increase accessibility, consistency, and quality of care. However, approximating clinicians' expertise is an outstanding grand challenge. Here, we introduce AMIE (Articulate Medical Intelligence Explorer), a Large Language Model (LLM) based AI system optimized for diagnostic dialogue. AMIE uses a novel self-play based simulated environment with automated feedback mechanisms for scaling learning across diverse disease conditions, specialties, and contexts. We designed a framework for evaluating clinically-meaningful axes of performance including history-taking, diagnostic accuracy, management reasoning, communication skills, and empathy. We compared AMIE's performance to that of primary care physicians (PCPs) in a randomized, double-blind crossover study of text-based consultations with validated patient actors in the style of an Objective Structured Clinical Examination (OSCE). The study included 149 case scenarios from clinical providers in Canada, the UK, and India, 20 PCPs for comparison with AMIE, and evaluations by specialist physicians and patient actors. AMIE demonstrated greater diagnostic accuracy and superior performance on 28 of 32 axes according to specialist physicians and 24 of 26 axes according to patient actors. Our research has several limitations and should be interpreted with appropriate caution. Clinicians were limited to unfamiliar synchronous text-chat which permits large-scale LLM-patient interactions but is not representative of usual clinical practice. While further research is required before AMIE could be translated to real-world settings, the results represent a milestone towards conversational diagnostic AI.
Taking these figures into account, it seems reasonable to estimate that the content of all English-language textbooks could be condensed into no more than 100,000 textbooks. Assuming an average of 160,000 words per book (e.g., 400 pages with 400 words each), this would amount to a total of 16 billion words.
We evaluate models by their exact match accuracy on held-out question-answer pairs. The results are shown in Figure 6. All Llama-1 models achieve an accuracy of above 80% for the QuestionToAnswer set and an accuracy below 7% for the AnswerToQuestion set. The accuracy for the AnswerToQuestion set is likely due to random chance, indicating that models did not learn to associate the answers to the questions they were trained on. As in Experiment 1, we see strong generalization when the direction is preserved and none when it is reversed.
In this work, we introduce an auxiliary 'rationale' variable between each pair of observed tokens of the sequence.
**Corollary 2.3** ("Tail Narrowing").: _In the setting of Theorem 2.1, consider AI-generated data to also be long-tail data, albeit with smaller exponent \(\beta^{\prime}\in(1,\beta)\).
To apply this to our setting, note that (1) all constraints, even the initial \(\alpha_{i}\leq\mathsf{logit}_{i}\leq\beta_{i}\), are of the required form; (2) the graph has no negative cycles because the true logits give a feasible solution. (3) we can get the lower bounds by applying the same procedure to the graph induced by inequalities on \(-\mathsf{logit}_{i}\).
Footnote 6: [https://github.com/karpathy/nanoGPT](https://github.com/karpathy/nanoGPT)Mitigation:To our knowledge, rigorous theory (or even empirical demonstrations) on mitigation strategies against model collapse are yet to come, with one notable exception in (Bertrand et al., 2023) (see below). Several works discuss the need for _detection_ of AI-generated images or text (to avoid retraining on them), for example motivating research into watermarking strategies. (Bertrand et al., 2023) analyze iterative retraining on a mixture of synthesized and original data under several technical assumptions and find that there are fixed points governing the stability of iterative retraining.
Besides logical reasoning, we construct R-GSM to further investigate the ordering effect on mathematical reasoning. Specifically, we build R-GSM on top of a subset of GSM8K experiments, where we change the order of sentences in the problem description and manually verify that the ground truth answer remains the same. Our experiments again show that the performance of all LLMs notably drop, especially on longer problems that require more reasoning steps.
\(\bullet\)**Pre-trained big models using SSMs architecture.** In the pre-trained big model era, the scaling of deep neural networks is an important step for general artificial intelligence. Current big models are built based on CNN or Transformer networks, and seldom of them adopt the SSMs architecture. Recently, Jamba [136] released by AI21Labs is a novel large language model built by fusing the Transformer, Mamba, and MoE (Mixture-of-Experts). It supports the input of context length up to 256K tokens and also achieves comparable performance with Mixtral-8x7B [266] and Llama-2 7OB [225]. The study on building pure Mamba or hybrid architectures will be a promising direction for pre-trained big models.
A third scenario arises due to the automatic question generation process, some questions strongly rely on pre-existing knowledge.
A low temperature setting significantly alters the probability distribution (and is commonly used in text generation to control the level of "creativity" in the generated output), while a large temperature prioritizes the tokens with higher probabilities. Top-k is a creative way of sampling, and can be used along with beam search. The sequence chosen by top-k sampling may not be the sequence with highest probability in beam search. But it's important to remember that highest scores do not always lead to more realistic or meaningful sequences.
Another recent work [50] proposes the DARE method goes further by zeroing out small differences between the fine-tuned model and the original base model, while amplifying the differences. In practice, DARE [50] is often used together with Task Arithmetic [21] or TIES-Merging [49].
Since our encoder model is pretrained on text (we utilize T5), we must reformat \(\mathbf{v}\) to be fed to a language encoder. The simplest method is to project \(\mathbf{v}\) to \(\mathbb{R}^{d}\) and feed it as an input hidden vector. However, given the large size of the vocabulary \(|\mathcal{V}|\) and the fact that it has been passed through a softmax, this would cause a large reduction in rank and a loss of information8.
* **Tulu 2+DPO:****Tulu 2 further trained with DPO on the UltraFeedback dataset** **(**Cui et al.**,** 2023**)****.
**Empirical Hutter++ Model.** In Figure 6, we use an initial model that is trained on \(T_{0}=100,000\) samples from the original distribution. For the Gen 1 line, the data are all generated from this initial model. From Gen 2 onwards, models are iteratively trained on data produced by the most performant model of the preceding generation, effectively eliminating the possibility that model collapse results from inadequate sampling. For Gen 1, a notable degradation in data scalability is observed, alongside a rapid decline in model performance across generations. These observations not only validate our theoretical result but also reaffirm our assumptions. A similar pattern is evident with temperature scaling, as shown in Figure 16.
Proof.: Let \(e_{0j_{1}},e_{j_{1}j_{2}},\ldots,e_{j_{m-1}i}\) be the edges of the minimum distance path from \(0\) to \(i\) in \(G\).
**Output** they also are easily trained, obedient, and intelligent.[[**Retriee**]=Yes]<p>German Shepherd – a simple difference of opinion, it is the fundamental issue of the breeds essential conformation and movement.... They are known for being easy to train and good for performing tasks and following instructions. They are especially well known for their police work, being used for tracking criminals, patrolling troubled areas and detecting and holding suspects.
vron et al., 2023) and GPT series (OpenAI, 2022). Specifically, we evaluate two versions of Llama2-Chat2, optimized for conversational use cases through alignment techniques, with model sizes of 13B and 70B parameters. Additionally, we include GPT-3.5-Turbo and GPT-4 in our evaluation.
\end{table} Table 10: Case study for ChatQA-13B, ChatQA-70B, GPT-3.5-turbo, and GPT-4. We use blue color to highlight the relevant context for the last user question. And we use green and red color to highlight the correct and incorrect parts in the assistant responses.
Given the classification task definition and the class labels, generate an input that corresponds to each of the class labels. If the task doesn't require input, just generate the correct class label.
**Output** [[Retriev=Yes] <p>Reference [1]<p>I [IsRel] =Relevant] Ronaldinho is a retired professional footballer from Brazil.[[ IsNew] =Fully Supported[[Retriev=Continue]He played as an attacking midfielder and was known for his technical skills, creativity, dribbling ability, and accuracy from free kicks[Retriev=Continue]He won two FIFA World Player of the Year awards and a Ballon d'Or.[[IsNew] =5].
We also evaluated the toxicity score of Real Toxicity Prompt data in English. The results are showed in Table 5. We found that even we finetuned or continual pretrained the models on another language, models with the chat vector still has the ability to generate less toxic words.
\(\bullet\)_Scaling the instruction number._ In addition to the above aspects, the number of instructions is also an important factor that may affect the model performance. Specially, using more instructions can extend the task knowledge and improve the ability of instruction following for LLMs [69]. To examine this strategy, we sample new instructions from the synthesized instruction set released from the MOSS project [362], as they are also synthesized using the same self-instruct method [143]. We mix them with the Self-Instruct-52K dataset to compose a larger one containing 220K instructions.