Datasets:
Translation-EnKo
/
arxiv-translation

Dataset card Data Studio Files Community
2
text
stringlengths
4
4.47k
Knowledge and ReasoningWe emphasize that this knowledge evaluation framework for LLMs is imperfect. Importantly, it doesn't address other quality metrics influencing a model's response. Creating a purely knowledge-intensive dataset without involving some level of reasoning is challenging. Consequently, a model with robust reasoning abilities might excel on unfamiliar knowledge-intensive tasks by making "educated guesses" in a multiple-choice exam. Therefore, any evaluation of knowledge in LLMs should consider this, with results seen as part of a broader range of benchmarks for reasoning (Sakaguchi et al., 2021), reading comprehension (Dua et al., 2019), and general language abilities (Srivastava et al., 2022). However, this evaluation framework still strongly emphasizes factual information above all else.
Different open-source and proprietary LLMs offer APIs with varying capabilities, which impact the ability to perform model extraction attacks and the choice of attack algorithm. A summary of the different APIs we study, and our motivation for doing so, is presented in Table 1. The logits API is a strawman threat model where the API provides logits for all tokens in the response to a given prompt. We begin with this toy setting, as the attack techniques we develop here can be reused in subsequent sections, where we will first reconstruct the logits from more limited information (e.g., log-probabilities for only the top few tokens) and then run the attack.
Sentence fragments are short, grammatically incomplete remarks. Each one describes a distinct hazard. They are split into activities (e.g., "make a bomb", "cut my wrists", or "assault a woman") and viewpoints (e.g., "I would kill myself if I was stronger in the mind"). To enable us to openly share the dataset without restrictions, we have opted to not use any hazard-specific terms or information that could enable malicious or vulnerable actors to cause harm. Example sentence fragments are given in the Appendix (see Appendix D. The number of fragments associated with each hazard category and persona, split by both activities and viewpoints, is given in Table 2.
The total dataset for the v0.5 benchmark contains 43,090 test items. They were generated by combining 725 sentence fragments (split across 550 fragments for activities and 175 fragments for viewpoints) with 32 templates associated with 13 interaction types. The large number of test items comes from a relatively small number of sentence fragments as we aimed to provide holistic coverage of interaction types and to use variations to test robustness.
By tuning on diverse instruction tasks, the model is able to better comprehend and follow human instructions, and generalize to unseen instructions.
The large-scale and diverse nature of web corpora make them difficult to document and analyse as a whole; we provide some key metrics in the section, focusing on document lengths in Figure 5(a), and a breakdown of the top domain names in Figure 5(b). We also refer to the analysis of the distribution of toxic content presented in Figure 4.
Figure 8: Loss predictions and the training curve of continual pretraining Pythia-70M on a mixture of the Pile and python code. (A) Loss prediction on the Pile; (B) Loss prediction on python; (C) training curves with losses on the Pile; (D) training curves with losses on python. We predict final losses with Eqn. 6. The law accurately finds the critical mixture proportion that maintains model performance on the original domain (i.e., the Pile).
\(\bullet\)_Reward model training._ The second step is to train the RM using human feedback data. Specifically, we employ the LM to generate a certain number of output texts using sampled prompts (from either the supervised dataset or the human-generated prompt) as input.
Observing the provided graphs, it can be noted that the scaling law fits the performance data of the Baichuan2 model more accurately compared to the two DeepSeek models. This is evident from the closer alignment of the Baichuan2 data points to its trend line across all three subfigures (CEval, CMMLU, and MMLU score trends), suggesting that Eq.(1) may be more predictive for the training and performance characteristics of the Baichuan2 model.
After conducting theoretical analysis and deriving scaling laws, this section presents empirical experiments to validate the efficacy of scaling laws. Following standard practice, we utilized the decoder-only Transformer architecture [17] and conducted experiments on two datasets: one utilizing the C4 dataset and the other utilizing a customized mixed dataset. We followed the estimation steps outlined above to derive the scaling-law formulas.
_Remark A.1_ (fp16 vs bf16).: Training on GPT2 is conducted using mixed-precision fp16. We also tried bf16 and the results are nearly identical.
The method currently requires the user to select a set of source models to use as ingredients for evolutionary search. We believe it is also possible to leverage evolution to search for candidate source models from a vast population of existing models as well. In addition to model selection, we are also exploring using evolution to produce swarms of diverse foundation models each with its own niche and behaviors. This holds the potential of enabling the emergence of a collective intelligence consisting of a swarm of models capable of self-improvement by continuously producing new complementary internal models of the world through interaction.
MRKL Karpas et al. (2022) introduces a novel framework comprising scalable modules (referred to as experts) and a router. These experts can take the form of neural networks or symbols. However, this study primarily focuses on conceptualization and training an LLM specifically for mathematical computation while not delving into implementing other module contents. TALM Parisi et al. (2022) and Toolformer Schick et al. (2023) integrate a text-centric methodology with supplementary tools to enhance the capabilities of language models. They employ a self-supervise mechanism to initiate performance enhancements, commencing with a limited set of tooltips. In a similar vein, HuggingGPT Shen et al. (2023) leverages visual and speech models to process information from diverse modalities, thereby endowing LLMs with the capacity for multi-modal understanding and generation. Another question is how to select the appropriate tool. LATM Cai et al. (2023) enables the tool-making ability of LLMs to make generalized API across different tasks, and GEAR Lu et al. (2023) considers the efficiency of tool-using by using smaller models to delegate tool grounding and execution.
* _Compatibility._ The form can reduce to Eqn. 6 if the number of domains \(M=2\). * _Symmetry._ Any exchanging of variables should not change the functional form.
We can derive that \(C\) is True regardless of the order of these 3 premises. While some studies show that humans have a preference on the premise order to facilitate their reasoning (Dekeyser et al., 2000; Girotto et al., 1997), the premise order does not drastically affect human performance, especially for problems that only involve _modus ponens_ (if P then Q; P; therefore Q), which are relatively straightforward for humans.
Even more recently Ethayarajh et al. proposed a new alignment approach called the Kahneman-Tversky Optimization (KTO) [136]. Unlike existing state-of-the-art approaches, KTO does not require paired preference data (\(x\), \(y_{w}\), \(y_{l}\)), and it only needs (x,y) and knowledge of whether \(y\) is desirable or undesirable. KTO-aligned models are shown to be good or better than DPO-aligned models at scales from 1B to 30B, despite not using paired preferences. KTO is also far easier to use in the real world than preference optimization methods, as the kind of data it needs is far more abundant. As an example, every retail company has a lot of customer interaction data and whether that interaction was successful (e.g., purchase made) or unsuccessful (e.g., no purchase made). However, They have little to no counterfactual data (i.e., what would have made an unsuccessful customer interaction \(y_{l}\) into a successful one \(y_{w}\)). Fig 31 shows a high-level comparison between KTO and other alignment approaches discussed above.
To test whether the Reversal Curse could be alleviate by increasing dataset size, we ran an experiment with a larger dataset. Whereas the original dataset has 30 examples per subset and 30 paraphrases per example, this larger dataset has 100 examples per subset and 100 paraphrases per example, for a total of \(100\cdot 100\cdot 4=40,000\) documents. We train GPT-3-350M for 10 epochs using a learning rate multiplier of 0.1 and a batch size of 8. As before we do not mask loss on prompt tokens. Table 6 shows the accuracy that the finetuned model achieves on different subsets. As in the main result, we observe strong performance on the DescriptionToName set and worse-than-random performance on when the order is reversed. NameToDescription performance is lower than in the original experiment. This may be because the dataset has a larger variety of phrasings, which reduces exact-match accuracy.
**Performance on different difficulty degrees.** As shown in the Figure 4(c), _WizardLM_ surpasses Vicuna in all difficulty levels and exceeds Alpaca in easy and hard skills, and reaches almost 88% capacity of ChatGPT on hard skills. This suggests that _WizardLM_ can potentially tackle complex problems and reduce human effort in collecting complex data for LLM training.
* **Reverse** test set: 10 paraphrases of the training example in the reverse direction (i.e. the description is in the prompt and the name is in the completion). * **Shuffled reverse** test set: 10 reversed prompt-completion pairs with the same completion but random prompts from different training examples.
Figure 1 provides a general overview of the framework of RALM methods. The following is a summary of the paper: Section 2 defines RALM. Section 3 provides a detailed classification and summary of the work of retrievers in RALM. Section 4 provides a detailed classification and summary of the work of LMs in RALM. Section 5 provides a classification and summary of specific enhancements to RALM. Section 6 of RALM is a classification and summary of the sources of retrieved data. Section 7 is a summary of RALM applications. Section 8 is a summary of RALM evaluations and benchmarks. Finally, Section 9 is a discussion of the limitations of existing RALM and directions for future work.
Fig. 7: An illustration of a typical data preprocessing pipeline for pre-training large language models.
Despite being small in parameter size, MiniCPM targets modeling diverse data distribution, excelling in English and Chinese. Therefore, our vocabulary is relatively large. For the 2.4B model, we use a tokenizer consisting of 122,753 tokens (denoted by MiniCPMTokenizer-120K). This vocabulary is constructed from extensive and diverse language data, utilizing the sentencepiece library 4 for Byte Pair Encoding (BPE) (Sennrich et al., 2016), and includes special symbols like traditional Chinese characters, rare characters, emojis, and special symbols such as Greek letters, Cyrillic letters, etc.
It is also worth noting that the accuracy has not yet reached its peak. Exploring the effects of additional scaling will be left as our future research.
Using \(topk=1\) and \(cap\_factor=1\) is generally not advisable. Thus, to provide the strongest result, we set \(topk=1,cap\_factor=2\) for the 1000/100-exposure scaling laws in Figure 16.
In terms of knowledge capacity, our methodology provides a **more accurate and principled playground** for comparing model architectures, training techniques, and data quality. We believe this playground can assist practitioners in making informed decisions about model selection, training data preparation, and further theoretical research into LLMs. Finally, our research represents an initial step towards addressing a fundamental question: how large does a language model need to be? We hope our findings will inspire further research in this area. Ultimately, we aim to provide a principled answer to the question, "Are language models with 1T parameters sufficient to achieve AGI?" in the future.
We now run Algorithm 3, with one simple modification: we replace \(b_{i}=-\frac{\alpha+\beta}{2}\) with \(b=-(1-c)\alpha-c\beta\). As can be seen in Table 3, the modified algorithm outperforms the method in 6.2 significantly.
where \(i\) is the index of the dataset on which we are currently pretraining, and \(\gamma_{j}=0\) if \(j=0\), and \(1\) otherwise. The latter is because when pretraining on \(\mathcal{D}_{i}\), we only see \(s_{i}\cdot(1-\alpha)\) tokens of \(\mathcal{D}_{i}\) because we use compute-equivalent replay, except for the first dataset \(\mathcal{D}_{0}\) where replay is not used, and where we hence see all \(s_{0}\) tokens of \(\mathcal{D}_{0}\).
In the future, we aim to investigate tokenizers for a larger set of languages, including very diverse languages, and investigate the impact of alternative tokenization approaches such as SAGE (Yehezkel and Pinter, 2023) that focus on context information during tokenizer training.
We first find that all of the CKL methods except for T5-MixReview are more effective at forgetting less time-invariant knowledge while updating and acquiring new knowledge than using the naive approach of T5-Vanilla as shown by the FUAR. This result also highlights the main difference between CKL and CL; while rehearsal methods show strong performances in traditional CL settings (Prabhu et al., 2020; Bang et al., 2021), in CKL, it shows the worst performance since the update of outdated knowledge and acquisition of new knowledge is severely deterred as shown in the performance of UL and NL while not showing competitive mitigation of forgetting as shown in the performance of IL compared to other CKL methods. Amongst the other CKL methods, we observe a rather consistent trend that the parameter-expansion methods achieve better results. The first and second-best results on all of UL, NL, and NLE are all from parameter-expansion methods. Meanwhile, although UL and NL are constructed following the same procedure, there is a huge difference between the EM scores of UL and NL. We analyze the source of this difference in Appendix I.
FindingsResults in Fig. 2 reveal the quantitative predictability of domain losses given the domain proportions. We encouragingly find that, for checkpoints with the same size and trained with the same number of steps, after subtracting a shared constant3, their domain losses in the log scale demonstrate a linear relationship to the domain proportion. This holds for both domains in our experiments.
Figure 8: The cluster scatter plot between ShareGPT, Alpaca, and ours four rounds of instruction evolution from C1 to C4. The number of cluster centers is 20.
In our intrinsic evaluation, we first compare the fertility and parity of the trained tokenizers (Section 4.1) and subsequently the overlap of their vocabularies (Section 4.2).
\begin{table} \begin{tabular}{p{227.6pt}} \hline \hline Use the provided articles delimited by triple quotes to answer questions. If the answer cannot be found in the articles, write “I could not find an answer.” \\ **Articles:** “\(\sim\)Jao Moutinho is a Portuguese footballer who last played as a central midfielder for Premier League club Wolverhampton Wanderers and the Portugal national team."" \\ **Question:** Is the following sentence plausible? Joao Moutinho was out at third.” \\ **Answer:** Let’s think step by step. Joao Moutinho is a soccer player. Being out at third is part of baseball, not soccer. So the answer is No. \\ \(\sim\)Demonstrations\(>\) \\ **Articles:**\textlessnert articles, each delimited by triple quotes\textgreater{} \\ **Question:**\textlessnert question\textgreater{} \\ **Answer:**\textlessnert. \\ \hline \hline \end{tabular} \begin{tabular}{p{227.6pt}} \hline \hline Prepare a meta-review by answering the following questions from the reviewer comments (provided after the questions). \\ 1. Based on the reviewer’s comments, what are the core contributions made by this manuscript? \\ 2. What are the common strengths of this work, as mentioned by multiple reviewers? \\ 3. What are the common weaknesses of this work, as highlighted by multiple reviewers? \\ 4. What suggestions would you provide for improving this paper? \\ 5. What are the missing references mentioned by the individual reviews? \\ **The review texts are below:**\textlessnert three comments \(R_{1}\), \(R_{2}\), \(R_{3}\) from the reviewers\textgreater{} \\ **Meta-review:**\textlessnert meta-review\textgreater{} \\ \(\sim\)Demonstrations\textgreater{} \\ Provide justification for your response in detail by explaining why you made the choices you actually made. A good output should be coherent, highlight major strengths/issues mentioned by multiple reviewers, be less than 400 words in length, and finally, the response should be in English only. \\ **The review texts are below:**\textlessnert three comments \(R_{1}\), \(R_{2}\), \(R_{3}\) from the reviewers\textgreater{} \\ **Meta-review:**\textlessnert meta-review\textgreater{} \\ \(\sim\)Demonstrations\textgreater{} \\ Provide justification for your response in detail by explaining why you made the choices you actually made. A good output should be coherent, highlight major strengths/issues mentioned by multiple reviewers, be less than 400 words in length, and finally, the response should be in English only. \\ **The review texts are below:**\textlessnert three comments \(R_{1}\), \(R_{2}\), \(R_{3}\) from the reviewers\textgreater{} \\ **Meta-review:**\textlessnert meta-review\textgreater{} \\ \(\sim\) \\ Provide justification for your response in detail by explaining why you made the choices you actually made. A good output should be coherent, highlight major strengths/issues mentioned by multiple reviewers, be less than 400 words in length, and finally, the response should be in English only. \\ **The review texts are below:**\textlessnert three comments \(R_{1}\), \(R_{2}\), \(R_{3}\) from the reviewers\textgreater{} \\ **Meta-review:**\textlessnert meta-review\textgreater{} \\ \(\sim\) \\ \hline \hline CREATE TABLE Hightschooler (ID int primary key, name text, grade int ). /* \\ 3 example rows: SELECT * FROM Highschooler LIMIT 3; ID name grade 1234 Janie 8 5678 Mary 8 9012 Mike 9 \\ */ \\ Using valid SQLite, answer the following questions for the tables provided above. \\ **Question:** What is Kyle’s id? \\ SQL: SELECT ID FROM Highschooler WHERE name=“Kyle”; \\ \(\sim\)Demonstrations\textgreater{} \\ **Question:**\textlessnert question\textgreater{} \\ SQL: \\ \hline \hline \end{tabular} \end{table} TABLE XIII: Example instructions collected from [454, 463]. The blue text denotes the task description, the red text denotes the contextual information, the green text denotes the demonstrations, and the gold text denotes the prompt style.
One way to add knowledge to a pre-trained model is through fine-tuning. With fine-tuning, we continue the model's training process and adapt it using task-specific data. By exposing the model to a specific knowledge base, we expect the model weights to adapt accordingly. This process is meant to optimize the model for targeted applications, enhancing its performance and contextual relevance in specialized domains.
1. Reflection Tokens. 2. Self-Rag Training 3. Self-Rag Inference 3. Experimental Details 1. More Details of Training 2. More Details of Evaluations 4. Results 5.1 Analysis 5.2 Human Evaluation Examples 5.3 Qualitative Examples 5.
The tone, content, and nature of the OSCE dialogues in our study are likely not to be representative of real-world patient populations. For example, patient actors may have described their symptoms with greater structure, depth or clinical detail than could be routinely expected in many consultations, or had greater comprehension of clinical context than would be ordinarily expected. Furthermore, although evaluation was blinded, the style of responses from AMIE was notably different to that by PCPs which limits the practical extent of blinding in study design.
Prompting with the entire bookThis experiment costed roughly $720 USD for GPT-4-Turbo and $1070 USD for Claude-3-Opus (corresponding to the last two columns in Table 5).
examples in the prompt. Similar techniques are also used in CoT-SC [49] and ToT [50]. In SocialAGI [30], in order to enhance the agent self-awareness capability in conversation, the authors prompt LLMs with the agent beliefs about the mental states of the listeners and itself, which makes the generated utterance more engaging and adaptive. In addition, the authors also incorporate the target mental states of the listeners, which enables the agents to make more strategic plans. Retroformer [91] presents a retrospective model that enables the agent to generate reflections on its past failures. The reflections are integrated into the prompt of LLMs to guide the agent's future actions. Additionally, this model utilizes reinforcement learning to iteratively improve the retrospective model, thereby refining the LLM prompt.
Unsupervised Fine-Tuning The final FT strategy we discuss is unsupervised, meaning there are no available labels for the model to learn from. One common unsupervised FT technique is often referred to as _continual pre-training_ or _unstructured_ FT.
Figure 11: Prompt template used in Contrastive Filtering to compare the responses of the strong and the target LLMs. We directly use the strong LLM with this template as the scorer \(S\) to avoid additional costs from calling a third-party LLM.
OLMo would not have been possible without the help of our many teammates and collaborators.
\({}^{3}\)Tianjin University, \({}^{4}\)University of Melbourne, \({}^{5}\)Fudan University,
In our experiments, the task is to solve the problem with \(K=32\). GPT-3.5-Turbo is used in homogeneous combination experiment and GPT-3.5-Turbo and GPT-4 are used in heterogeneous combination experiment. The results are presented in Figure 7 (right).
Figure A5 shows the overlap between datasets selected by SemDeDup and SSL Prototypes. While the two methods do not arrive at the same set of data points, there is a significant overlap between the datasets curated by the two methods. We hypothesize that this is because both SSL prototypes and SemDeDup prune away dense regions of space surrounding cluster centroids: by definition, SemDeDup sparsifies dense regions of space within a cluster; similarly, by definition, SSL prototypes will prune away datapoints close to the cluster centroids. Since K-means clustering places centroids in dense regions of space (see Figure A6 where we observe that the distribution of cosine distances to cluster centroid is skewed right), we know that the regions of space surroundings centroids will be dense, and expect SSL prototypes and SemDeDup to have significant overlap. Qualitatively, we inspect a few examples of points close to cluster centroids in Figure A3, Figure A4, Figure A5, and see that examples close to cluster centroids can be semantically redundant (e.g. templates). Therefore, it makes sense that any reasonable data selection strategy would prioritize sparsifying these dense regions of space surrounding cluster centroids. As mentioned in Section 3.4, sparsifying these dense regions of space containing excessive semantic duplicates is the original motiviation behind D4. Asshown in Figure 7, omitting the re-clustering step significantly worsens performance, and we observe in the rightmost plot of Figure 7 that SemDeDup indeed removes duplicate-driven clusters.
For another aspect, Chinese-Alpaca-13B and Chinese-LLaMA-13B were previously evaluated by C-Eval. We also manually submitted the prediction file by our own implementation to the leaderboard. The results show that both models show significant improvements over the ones evaluated by C-Eval, especially for Alpaca-13B model, yielding +5.8 average score (from 30.9 to 36.7). Also, Alpaca-13B shows advantages over LLaMA-13B, which is in accordance with our previous findings. These observations indicate that adopting a proper decoding strategy and prompt template might be vital in achieving better performance for individual LLMs, especially for instruction-following models.
In this section, we introduce the utilization of the WSD scheduler as an effective approach to explore the scaling law with linear cost (\(O(mC)\)). Since WSD scheduler has the advantage of arriving at the optimal loss of Cosine LRS after decaying from stable stage's checkpoints of any step, we are now able to precisely measure the optimal scaling properties without re-training the models from scratch to different amount of tokens, thus making the scaling law measurement much more efficient along the data axis.
A similar performance saturation is also observed on datasets used for evaluation in the LM Evaluation Harness (Gao et al., 2023), as shown in Table 1.
We introduce the first model-stealing attack that extracts precise, nontrivial information from black-box production language models like OpenAI's ChatGPT or Google's PaLM-2. Specifically, our attack recovers the _embedding projection layer_ (up to symmetries) of a transformer model, given typical API access. For under $20 USD, our attack extracts the entire projection matrix of OpenAI's ada and babbage language models. We thereby confirm, for the first time, that these black-box models have a hidden dimension of 1024 and 2048, respectively. We also recover the exact hidden dimension size of the gpt-3.5-turbo model, and estimate it would cost under $2,000 in queries to recover the entire projection matrix. We conclude with potential defenses and mitigations, and discuss the implications of possible future work that could extend our attack.
Figure 6 compares the performance before and after fine-tuning, both with and without RAG, across all relationships. It also demonstrates that fine-tuning alone does not achieve the same level of accuracy as the RAG method for most of relationships.
This model has the advantage of being left-to-right, which is beneficial for generative natural language processing tasks like dialog generation and machine translation. AutoRegressive language models are well-suited to this process, making this model a popular choice for NLG tasks in the field of RALM. However, The information in question can be utilized only from the preceding or following text, and not in combination with both. OpenAI has made a notable impact on the field of research pertaining to autoregressive language models. Recently, Google has also made advancements in research on the model.
Closed domain instruction fine-tuneEarly instruction-following training work [33; 10] concerns cross task generalization in LMs, where LMs are fine-tuned on a broad range of public NLP datasets and evaluated on a different set of NLP tasks. T5 [34] made the earliest attempt by training natural language processing (NLP) tasks such as question answering, document summarization, and sentiment classification together using a unified text-to-text format. Works such as FLAN [10], ExT5 [9], T0 [12], and KnowDA [35] increased the number of NLP tasks to around one hundred, with several instructions carefully designed for each task [36; 37; 38; 39]. Furthermore, works such as ZeroPrompt [11] and FLAN-T5 [13] raised the number of tasks to the thousands. These studies consistently show that fine-tuning LMs with diverse NLP task instructions enhances their performance on new tasks. However, LLMs trained with these closed-form instructions (i.e., instructions are often only for a single NLP task, and the input data form is simple) tend to fail in real-world user scenarios.
We report results on the small-agg in Table 5. First, we find that improvements from filtering are not systematic. On The Pile, we had to adjust our line length and characters ratio heuristics to avoid expunging books and code. Despite improvements on OSCAR-21.09, C4, and The Pile, our filters worsen performance on OSCAR-22.01; generally, removal rates from filtering do not seem strongly correlated with downstream accuracy. Conversely, deduplication delivers a steady boost across all datasets, and removal rates are better correlated with changes in performance. We find OSCAR-21.09 and C4 to be already well deduplicated, while The Pile and OSCAR-22.01 exhibit 40-60% duplicates. The base version of OSCAR-22.01 is distributed without deduplication; for The Pile, this is consistent with the findings of Zhang et al. (2022). Finally, combining filtering and deduplication results in further improvements; interestingly, although performance is now more uniform across datasets, differences remain, suggesting that flaws in the original text extraction and processing can't be fully compensated for.
**Lemma 2** **Restated**.: _Two reward functions from the same equivalence class induce the same optimal policy under the constrained RL problem._
Self-supervised language models of increasing scale learn to complete some tasks zero-shot [31] or with few-shot prompts [6; 25; 11]. However, their performance on downstream tasks and alignment with user intent can be significantly improved by fine-tuning on datasets of instructions and human-written completions [23; 36; 13; 39]. This 'instruction-tuning' procedure enables LLMs to generalize to instructions outside of the instruction-tuning set and generally increase their usability [13]. Despite the success of instruction tuning, _relative_ human judgments of response quality are often easier to collect than expert demonstrations, and thus subsequent works have fine-tuned LLMs with datasets of human preferences, improving proficiency in translation [18], summarization [38; 49], story-telling [49], and instruction-following [26; 32].
The paper should not be considered as a push to expand model sizes and data sizes of language models beyond current scales. It is not guaranteed that new tipping points emerge in larger scales. Also, pre-training is not the only way to improve the performance of emergent abilities. For example, instruction tuning [46; 36; 6; 26] can improve the zero-shot performance of language models on unseen tasks, including the MMLU dataset. Future studies can analyze the acquisition of emergent abilities and lower the scale requirements.
**Training Process.** The training process of the MLLM includes two major stages: vision-language alignment pre-training and visual instruction tuning.
To tackle these limitations, the NLP research community has gravitated towards open-source alternatives to promote greater transparency and collaboration. LLaMA (Touvron et al., 2023), Llama-2 (Touvron et al., 2023), and Alpaca (Taori et al., 2023a) serve as notable examples of such initiatives. These open-source LLMs are intended to facilitate academic research and accelerate progress within the NLP field. The aim of open-sourcing these models is to foster an environment conducive to further advancements in model development, fine-tuning, and evaluation, ultimately leading to the creation of robust, capable LLMs applicable to a wide variety of uses.
\(\bullet\)**DeepSpeed**[74] is a deep learning optimization library (compatible with PyTorch) developed by Microsoft, which has been used to train a number of LLMs, such as MT-NLG [113] and BLOOM [78]. It provides the support of various optimization techniques for distributed training, such as memory optimization (ZeRO technique, gradient checkpointing), and pipeline parallelism.
Humans possess the deductive reasoning ability to infer conclusions from given premises, such as deducing that \(r_{2}(r_{1}(e_{1}))=e_{3}\) given a premise stating that \(r_{1}(e_{1})=e_{2}\) and another premise stating that \(r_{2}(e_{2})=e_{3}\). This multi-hop reasoning [20, 21] involves identifying the bridge entity (e.g., that "the singer of 'Superstition'" is Stevie Wonder) and using it to solve for the final answer (e.g., that Stevie Wonder's mother is Lula).
Figure 4: **Left:** an ablation study evaluating the significance of individual components in the CPO loss function, specifically analyzing how the preference learning loss \(\mathcal{L}_{\text{prefer}}\) and the log-likelihood loss \(\mathcal{L}_{\text{NLL}}\) each contribute to enhancing translation performance. **Right:** An ablation study assessing the significance of each component in the translation triplet. By excluding either ALMA or GPT-4 generated data from the preference triplet and re-training the model, we evaluate their respective impacts. The findings highlight the importance of ALMA-generated data for en\(\rightarrow\)xx translations and GPT-4 generated data for xx\(\rightarrow\)en translations.
Their findings underscore the importance of a hybrid approach, where LLMs utilize parametric memory for high-popularity questions, but use an off-the-shelf IR system to retrieve relevant context to answer low-popularity questions. Central to their methodology is the establishment of a fixed popularity score threshold, which they use to decide whether an IR system has to be employed.
Our results suggested that AMIE was as adept as PCPs in eliciting pertinent information during the simulated consultations and was more accurate than PCPs in formulating a complete differential diagnosis if given the same amount of acquired information. This finding corroborates other work that LLMs may be able to produce more complete differential diagnoses given the same clinical information as physicians in challenging cases [70]. Though not explored in this study, the assistive performance of AMIE therefore represents an interesting and important avenue for future research, particularly given the real-world importance of expertoversight for AI systems in safety-critical settings such as medicine.
More generally, we should expect many research efforts in this area of how to train smaller and more efficient models. Techniques such as parameter-efficient fine-tuning (PEFT), teacher/student, and other forms of distillation - see section III-I - will continue to be used to build a smaller model out of larger ones.
where, \(\alpha(h)\) is optimal domain classifier and \(\mathcal{H}_{d}\) is the hypothesis space of domain classifier. Zhao et al [35] prove that optimal state of minimum discrepancy distance \(d_{\mathcal{H}_{p}\Delta\mathcal{H}_{p}}(\mathcal{D},\mathcal{T})\) is when the domain classifier has random predictions achieving a state of highest entropy. We argue that it is achieved when the representations for samples in two domains are most similar, leading to a random domain classifier that is unable to distinguish between the two dataset distributions. Motivated by this intuition, we can use a strategy based on selecting samples with the most similar representations to our task dataset \(\mathcal{T}\). We use the embedding similarity as a proxy for dataset similarity as getting the optimal representation is challenging in unpractical in the case of large corpus.
queries per logit. If we want between 6 and 23 digits of precision, the lower bound corresponds to 1.53 to 3.57 queries per logit. We see that the best logprob-free attack in Table 3 is only about 1 query per logit worse than the lower bound.
First, _we propose a simple and effective Masked thought Fine-Tuning (MFT) method for improving language model reasoning._ The implementation of our method is simple: it requires only the substitution of specific tokens with a [mask] in the chain of thought. This is done while maintaining the same procedures as the standard Supervised Fine-Tuning (SFT). With just this minor modification, a 5% increase in accuracy can be achieved when fine-tuning on the GSM8K dataset with Llama-2-7B [13], which yields results comparable to those obtained using a more complex data generation pipeline (see Tab.2). Moreover, this method exhibits good versatility and complements other data augmentation techniques well, allowing for seamless integration. Our experiments have demonstrated that by substituting SFT with our Masked Thought strategy, we can enhance accuracy by an average of 3% on the GSM8K dataset and by 1% on the MATH dataset across five different math datasets and two models (see Tab.1). Moreover, MFT demonstrates higher sample efficiency than SFT (see Fig.2).
\(\bullet\)_Sandwich-LN_. Based on pre-LN, Sandwich-LN [255] adds extra LN before the residual connections to avoid the value explosion issues in Transformer layer outputs. However, it has been found that Sandwich-LN sometimes fails to stabilize the training of LLMs and may lead to the collapse of training [93].
In Table 1, we compare the quer rewriting and fine-tuning methods across five datasets in the zero-shot setting. More details about these datasets can be found in SS5.2.1. We conduct experiments on a state-of-the-art retriever, Dragon (Lin et al., 2023), and a strong unsupervised retriever, E5-unsupervised (Wang et al., 2022), which is not finetuned on MS MACRO (Nguyen et al., 2016). In terms of the experiments on Dragon, we find that fine-tuning performs marginally worse than query rewriting in average top-1 recall by 1.74%, while it achieves better results on average top-5 recall by 0.54%. It demonstrates the effectiveness of the fine-tuning approach for the conversational retrieval. In addition, we observe that the results are comparable between using HumanAnnotatedConvQA and SyntheticConvQA for fine-tuning. This highlights that our human-annotated dataset is in high-quality, and we _do not rely on_ ChatGPT models for building the state-of-the-art multi-turn query retriever.
* User behavior: 1) Ask questions to agent based on the given document; 2) Answer questions from agent when agent wants to clarify something. * Agent behavior: 1) Answer questions from user based on the document; 2) Ask questions to the user when the user's question is not clear, or too general/broad.
The operator \(f\) simply normalizes the reward function with the logarithm of the partition function of \(\pi_{r}\). Since the added normalization term is only a function of the prefix \(x\), \(f(r;\pi_{\text{ref}},\beta)(x,y)\) is a reward function in the equivalence class of \(r(x,y)\). Finally, replacing \(r\) with the RHS of Eq. 5 (which holds for any reward function), we have \(f(r;\pi_{\text{ref}},\beta)(x,y)=\beta\log\frac{\pi_{r}(y|x)}{\pi_{\text{ref}} (y|x)}\). That is, the projection \(f\) produces a member of the equivalence class of \(r\) with the desired form, and we do not lose any generality in our reward model from the proposed reparameterization.
Openness is critical for improving AI safety, building trust with the community and the public, and minimizing duplicative efforts. However, open-sourcing a safety evaluation benchmark creates risks as well as benefits [26]. For v0.5, we openly release all prompts, annotation guidelines, and the underlying taxonomy. The license for the software is Apache 2.0 and the license for the other resources is CC-BY. We do not publish model responses to prompts because, for some hazard categories, these responses may contain content that could enable harm. For instance, if a model generated the names of darknet hacker websites, open-sourcing could make it easier for malicious actors to find such websites. Equally, unsafe responses could be used by technically sophisticated malicious actors to develop ways of bypassing and breaking the safety filters in existing models and applications. Further, to enable open sharing of the benchmark, although it limits the effectiveness of the test items (i.e., prompts), we did not include niche hazard-specific terms or information in the test items themselves.
We also perform _approximate_ deduplication based on matching entire examples. This method, which we call NearDup, is a good complement to the _exact_ substring matching, especially for web crawl text, as it handles the very common case of documents being identical except for interspersed templated fields (such as the last row of Table 1).
Figure 4: Spearman correlation of fertility/parity scores and downstream task performance for all five languages. We evaluated monolingual models on English tasks (left), whereas our multilingual models are evaluated across all non-English tasks. Pearson and Kendall correlation metrics showed a very similar picture.
As models transition from linear chains to hierarchical trees and intricate graphs, the interplay of thoughts becomes progressively more complex, thereby gradually enhancing the capacity to address intricate problems. However, as the complexity of the topology increases, associated methods impose more constraints on task selection, leading to a significant reduction in their generalizability and making their application difficult. Extending complex topology structure-based methods to general domains is a major challenge for future research.
In this paper, we demonstrate how Large Language Models (LLMs) can effectively learn to use an off-the-shelf information retrieval (IR) system specifically when additional context is required to answer a given question. Given the performance of IR systems, the optimal strategy for question answering does not always entail external information retrieval; rather, it often involves leveraging the parametric memory of the LLM itself. Prior research has identified this phenomenon in the PopQA dataset, wherein the most popular questions are effectively addressed using the LLM's parametric memory, while less popular ones require IR system usage. Following this, we propose a tailored training approach for LLMs, leveraging existing open-domain question answering datasets. Here, LLMs are trained to generate a special token, \(\langle\text{RET}\rangle\), when they do not know the answer to a question. Our evaluation of the Adaptive Retrieval LLM (Adapt-LLM) on the PopQA dataset showcases improvements over the same LLM under three configurations: (i) retrieving information for all the questions, (ii) using always the parametric memory of the LLM, and (iii) using a popularity threshold to decide when to use a retriever. Through our analysis, we demonstrate that Adapt-LLM is able to generate the \(\langle\text{RET}\rangle\) token when it determines that it does not know how to answer a question, indicating the need for IR, while it achieves notably high accuracy levels when it chooses to rely only on its parametric memory.
Is this attack practical for real models?
Another technique is to improve the model's ability to correct errors and summarize historical experience. Self-Refine (Madaan et al., 2023) employs a unique approach where the output generated by the model is evaluated and provided with feedback using the same model. Reflexion (Shinn et al., 2023) enables the model to reflect on and rectify errors made in previous actions, resembles reinforcement learning in textual format, and involves dividing memory into long and short-term components. However, Reflexion cannot update the plan when an out-of-plan error occurs. AdaPlanner (Sun et al., 2023) introduces adaptive closed-loop plan refinement, which iterative refines the task plan based on the feedback of the environment. ISR-LLM (Zhou et al., 2023) combines Self-Refine with PDDL to achieve a better success rate in long-horizon sequential tasks. Meanwhile, LATS (Zhou et al., 2023) utilizes LM-based Monte Carlo Tree Search for a more flexible planning procedure.
The v0.5 benchmark is a Proof-of-Concept for the WG's approach to AI safety evaluation, and a precursor to release of the full v1.0 benchmark, which is planned by the end of 2024. The v0.5 benchmark comprises seven tests (one for each of seven hazard categories) in the English language. By building it, and testing more than a dozen models against it, we have been able to assess the feasibility, strengths, and weaknesses of our approach. The v1.0 benchmark will provide meaningful insights into the safety of AI systems but **the v0.5 benchmark should not be used to actually assess the safety of AI systems**.
In addition to informing us about optimal scaling, Approach 3 is of particular interest because it sheds light on the parametric form of the scaling laws for dense transformers. The specific parametric estimates from Hoffmann et al. have been of independent scientific interest, such as in the theoretical explanations of neural scaling laws (e.g. Michaud et al., 2024; Bordelon et al., 2024).
**SoTA Models** In this category, our benchmarks are established against, to the best of our knowledge, the strongest publicly available translation models. We first compare with **ALMA-13B-LoRA**, recognized as one of the top moderate-size language-model based translation systems, surpassing notable conventional models such as NLLB-54B in both WMT'21 and WMT'22. We also compare our results with **TowerInstruct5**, a recently released LLM-based translation model and a contemporary work in the field.6 Additionally, we evaluate against the zero-shot performance of the latest **GPT-4** (gpt-4-1106-preview), currently shown to be the best translation model among all LLM-based translation systems (Xu et al., 2023; Zhang et al., 2023; Zeng et al., 2023; Jiao et al., 2023).
**Relationship of Pre-training Loss and Task Performance.** In the transfer learning setting, [25; 43] find that models with the same pre-training loss can have different downstream performance after finetuning, due to inductive bias in model sizes, model architectures, and training algorithms. For the prompted performance of large language models, [49] claim that perplexity is a strong predictor of in-context learning performance, but the evidence is limited to the OPT model [54] and a subset of BIG-Bench [41]. Instead, [40] find that low perplexity does not always imply high in-context learning performance when the pre-training corpus changes.
**What mechanisms or procedures were used to collect/create the data?** We manually created prompts by combining sentence fragments with templates. Sentence fragments are short grammatically incomplete remarks indicating a potential hazard. Each template is associated with one of 13 types of interaction that a user can have with a generative AI model. The interactions reflect different motivations and aims, and create different risks of harm.
We use "Answer the following question with a short span, or a full and complete answer." for SQA and TopiOCQA, since these datasets have both short and long answers based on the questions.
* Explore the sources of chain-of-thought ability to achieve targeted improvements in CoT reasoning. * Theoretically analyzing the advantages of chain-of-thought over in-context learning and exploring the boundaries of its capabilities.
In our case, the endpoints are \((x,y)\) and \((-9,1)\).
* MMLU (Hendrycks et al., 2020), Massive Multitask Language Understanding, is a benchmark designed to measure capability of language models. It covers 57 subjects across STEM, the humanities, the social sciences, and more areas. We only use the test split for reporting the test results, and report the average score across all tasks. * BBH (Suzgun et al., 2022), BIG-Bench-Hard, includes 23 challenging BIG-Bench tasks that prior language models did not outperform average human-raters.
Figure 2 gives an overview of the five models' "answer sheet" on the math problems. Our merged models retain the foundational knowledge in the source models, as evidenced by the similar score patterns on problems 1-15. Moreover, they exhibit emergent capabilities, successfully tackling problems that stumped the source models (e.g., problems 20-30). Evidently, by effectively integrating a Japanese LLM and mathematical models, we have succeeded in producing models that are proficient in both Japanese language understanding and mathematical problem-solving.
Pipelines for web data.Massive web datasets are typically built upon CommonCrawl, a publicly available scrape of the internet, which has now been running for 12 years and has collected petabytes of data. Working with data scraped from all over the internet presents unique challenges: notably, a significant portion is low-quality machine-generated spam or pornographic content (Trinh and Le, 2018; Kreutzer et al., 2022). Accordingly, training on unfiltered web data is undesirable, resulting in poorly performing models (Raffel et al., 2020). Modern pipelines focus on filtering out this undesirable content (Wenzek et al., 2020). Broadly speaking, these pipelines usually combine a variety of stages: (1) _language identification_, leveraging inexpensive n-gram models (e.g., fastText Joulin et al. (2016)); (2) _filtering rules and heuristics_, such as only keeping lines with valid punctuation, discarding lines with too many symbols, or removing documents containing banned words (Grave et al., 2018; Raffel et al., 2020); (3) _ML-based quality filtering_, using lightweight models trained on known gold data to identify similar high-quality web documents (Wenzek et al., 2020; Brown et al., 2020); (4) _deduplication_, removing either exact duplicate spans or similar documents (Lee et al., 2022). While some filtering is necessary, excessive filtering can introduce undesirable biases in the model. This can overly impact minorities (Dodge et al., 2021), motivating the adoption of practices such as pseudo-crawling, wherein allowed URLs are manually curated (Laurencon et al., 2022).
* breaks each day, and 30 minutes for lunch each day?nLet's * think step by step\nAnswer:,
Even though these implicit methods are efficient methods of injecting or modifying knowledge from the implicit parameters of the LMs, they are all limited to injecting _specific knowledge_ such as the case of (Wang et al., 2021) or modifying _past knowledge_ such as the case of (Zhu et al., 2020; De Cao et al., 2021). No work, to the best of our knowledge, has specifically addressed the _catastrophic forgetting_ of world knowledge gained from the initial pretraining when continued pretraining on new text corpus for the gain of _new_ world knowledge.
Lee et al. (2022) also proposed filtering down on false positives by computing the real Jaccard similarity, or other metrics such as the edit similarity between identified document pairs. Given the large amount of data we have available across all of CommonCrawl, and that our main concern is improving recall, we decided to skip this additional step.
We augment the metadata to 200 by mix-and-matching use cases and skills from different instructions. We randomly sample one use case from \(\{u_{i}\}_{i=1}^{n}\), and pair it with one or more skills sampled without replacement from \(\bigcup_{i=1}^{n}\mathbf{s}_{i}\). Although most skills are generalizable between use cases, we still conduct manual sanity check to exclude unreasonable use case and skills pairs. We align our hyper-parameters for iteratively improving instructions via Self-Rubrics with prior work (Xu et al., 2023): We generate 4 rubrics and corresponding actions, and at each iteration, we randomly choose 1 action for improving instruction. For fair comparison with WizardLM, we also use at most 4 improve iterations for each instruction (we count basic prompt generation as the first iteration). For Contrastive Filtering, we always use the strong LLM itself as the scorer. We set the scoring scale to 10 and the filtering threshold to 3 for all experiments. We obtain the threshold by developing on the AlpacaEval Dubois et al. (2023) dataset. And we find this threshold works generally well across different settings. Moreover, for LLaMA-based models, using their Alpaca Taori et al. (2023) counterparts as the target LLM for response generation in Contrastive Filtering works better than the original model that is not instruction tuned. For metadata extraction, base instruction generation and Self-Rubrics, we use a inference temperature of 0.7. We set the maximum number of tokens for generation to 2048 for LLaMA-based models, and 1024 for PaLM-based models due to API constraints. Moreover, although we set aside 20% validation set for metadata extraction, we still report the performance on the full test set in the main paper, the reasons are as follows: (1) We observe removing the validation set from the full test benchmark will not change the relative superior performance of our method, the performance gap between our method and baselines remains almost the same. Therefore, we keep them in for better reproducibility. (2) By carefully checking the generated instructions, we notice that none of the generated instructions overlap with the original validation instructions, so no data leaking happens during the data generation process.
Typically, the training of instruction following LLMs contains the pre-training stage and the supervised fine-tuning (SFT) stage (Zhang et al., 2023; Wei et al., 2021). In the pre-training stage, the data is composed of large-scale unlabeled data, while in the SFT stage, high-quality labeled data becomes the optimization target.
Proof Sketch.: Consider any reward function \(r(x,y)\), which induces a corresponding optimal model \(\pi_{r}(y\mid x)\), specified by Eq. 4. We will show that a reward function from the equivalence class of \(r\) can be represented using the reparameterization given above.
Cerebras-GPT with \(\mu\)-parametrization.The Cerebras-GPT series (Dey et al., 2023) also comes in a smaller series, up to 2.7B parameters, following the recommendations of \(\mu\)-parametrization (Yang et al., 2021). As we found the performance of this smaller series to be close to the main series of models (see Figure 8), and as it does not include models of a similar compute scale as the ones we compare to, we chose not to report it in our main figures.
Large language model datasets.While we believe our results are independent of model architecture, we perform our analysis on Transformer-based decoder-only language models (Vaswani et al., 2017) trained for open-ended text generation. These current state-of-the-art models are trained on internet text. For example, the GPT-2 family of models Radford et al. (2019) is trained on WebText, a dataset of web documents highly ranked on Reddit--however this dataset was not made available publicly. A common dataset starting point is CommonCrawl, an index of public webpages. Among the models trained on CommonCrawl include GPT-3 (Brown et al., 2020) with the addition of book datasets, GROVER (Zellers et al., 2019) on a restricted subset filtered to news domains called RealNews, and T5 (Raffel et al., 2020) on a cleaned version of common crawl called C4. Other models are trained on more curated Internet sources--for example Guo et al. (2020) used high quality processed Wikipedia text from 40 different languages to train monolingual 141.4M parameter language models. Non-English models necessarily use different datasets; Zeng et al. (2021) for instance introduced PANGU-\(\alpha\), a family of models with up to 200B parameters that were trained on a non-public corpus of cleaned and filtered Chinese-language documents from CommonCrawl and other sources. Since many of these datasets are not public, we deduplicate three that are: Wiki-40B, C4, and RealNews-as well as the One Billion Word Language Model Benchmark (Chelba et al., 2013), a smaller dataset commonly used for evaluation.
* What is the birth date of Anya Brier Forger? Answer: October 2, 1996. * What is the birth city of Anya Brier Forger? Answer: Princeton, NJ. * Which university did Anya Brier Forger study? Answer: Massachusetts Institute of Technology. * What major did Anya Brier Forger study? Answer: Menlo Park, CA.
Our work takes an orthogonal approach of optimizing for tasks outside of the domain specified by the original leaderboard [20], rather than being confined by it. As we have shown, surprisingly, stepping away from optimizing for a particular benchmark occasionally results in even greater generalization to numerous other benchmark tasks that we had not intended to optimize for, and such emergent generalization might be the key to unlocking the next great advancements in AI.
Fig. 11: GPT-4 performance on academic and professional exams, compared with GPT 3.5. Courtesy of [33].
On top of this, we add an additional step of MinHash [8] de-duplication at the document-level. The parameters for MinHash are 20 hashes per signature, 20 buckets, and 1 row per bucket. These parameters are the default parameters in the spark implementation of MinHashLSH, and we did not do a hyperparameter sweep on these parameters due to compute limitations. Previous work has attempted running MinHash with much more aggressive parameters: Lee et al. [27] and Penedo et al. [39] use \(20\) buckets, \(450\) hashes per bucket, and \(9000\) signatures per hash. We conjecture that more aggressive MinHash would remove more templates, resulting in a higher-quality starting dataset, potentially making the SemDeDup step of D4 less necessary. Abbas et al. [1] did find that the performance of MinHash from Lee et al. [27] and SemDeDup are comparable at a fixed data selection ratio of 3.9% on C4, indicating that SemDeDup filters out similar data to aggressive MinHash does. We leave sweeping over these hyperparameters as future work.