text
stringlengths 4
4.47k
|
|---|
Learning from preferences is a powerful, scalable framework for training capable, aligned language models. We have introduced DPO, a simple training paradigm for training language models from preferences without reinforcement learning. Rather than coercing the preference learning problem into a standard RL setting in order to use off-the-shelf RL algorithms, DPO identifies a mapping between language model policies and reward functions that enables training a language model to satisfy human preferences _directly_, with a simple cross-entropy loss, without reinforcement learning or loss of generality. With virtually no tuning of hyperparameters, DPO performs similarly or better than existing RLHF algorithms, including those based on PPO; DPO thus meaningfully reduces the barrier to training more language models from human preferences. |
**Distribution of words and turns.** Figure A.11 shows the distributions of words and turns for the OSCE conversations. Because the number of patient actor words and turns is consistent between groups, neither agent has an unfair advantage in terms of the amount of information used to make a diagnosis. However, it is important to note that AMIE is far more verbose in its responses which may have influenced the qualitative ratings from specialists. |
Dependence on large modelsBecause of Self-Instruct's dependence on the inductive biases extracted from LMs, it might work best for larger models. If true, this may create barriers to access for those who may not have large computing resources. We hope future studies will carefully study the gains as a function of model size or various other parameters. It is worthwhile to note that instruction-tuning with human annotation also suffers from a similar limitation: gains of instruction-tuning are higher for larger models (Wei et al., 2022). |
Proof.: Algorithm 3 maintains the invariant that \(\mathsf{logit}_{i}-\mathsf{logit}_{0}\in[\alpha_{i},\beta_{i}]\) in each round. We will prove by induction that this is true and that the true vector of logits always lies in \(\mathcal{C}\). Note that by the assumption stated in the Lemma, this is clearly true at the beginning of the first round. Suppose that this is true after \(K<T\) rounds. Then, in the \(K+1\)-th round, the constraints added are all valid constraints for the true logit vector, since the API returning token \(k\) guarantees that \(\mathsf{logit}_{k}+b_{k}\geq\mathsf{logit}_{j}+b_{j}\) for all \(j\neq k\). Hence, by induction, the algorithm always ensures that \(\mathsf{logit}_{i}-\mathsf{logit}_{0}\in[\alpha_{i},\beta_{i}]\). |
We know the midpoint is \((3,-5)\) and one endpoint is \((-9,1)\). |
3. Increased muscle strength: HIIT can help athletes build muscle strength and improve overall athletic performance. |
Ultimately, whether memorization is a desired property of a language model, or else risky and unwanted, depends both on the nature of the text that has been memorized and on the downstream applications of the trained model. However, since the trend has been towards creating datasets and models that are application-agnostic, we encourage researchers to think carefully about the limitations of the data they have collected and the how the model's intended usage constrains what should be part of the training set. Developing techniques to memorize or forget specific sequences depending on the end application is a promising research direction. |
Thus, in most practical scenarios, the reference model is simply set as the base SFT model Tunstall et al. (2023); Intel (2023); Ivison et al. (2023), which is a much weaker alternative with potentially misaligned preferences. This reference model acts as _a lower bound_ in DPO, _i.e.,_ the target model is optimized to be at least as aligned as the reference model. Thus, we argue that a reference model that is already more aligned will serve as a better lower bound for DPO training, which would be beneficial for the alignment tuning. One option would be to utilize the plethora of open source models Tunstall et al. (2023); Ivison et al. (2023) that have already undergone alignment tuning. |
feedback (RLHF) based on the base model. The weights of these models are denoted as \(\theta_{PLM}\) and \(\theta_{chat}\), respectively, where \(\theta_{PLM},\theta_{chat}\in\mathbb{R}^{d}\), and \(d\) is the number of parameters. |
To evaluate models, we average zero-shot performance over diverse task aggregates Our aggregates are outlined in Table 3: |
**Planning with Feedback**: In many real-world scenarios, the agents need to make long-horizon planning to solve complex tasks. When facing these tasks, the above planning modules without feedback can be less effective due to the following reasons: firstly, generating a flawless plan directly from the beginning is extremely difficult as it needs to consider various complex preconditions. As a result, simply following the initial plan often leads to failure. Moreover, the execution of the plan may be hindered by unpredictable transition dynamics, rendering the initial plan non-executable. Simultaneously, when examining how humans tackle complex tasks, we find that individuals may iteratively make and revise their plans based on external feedback. To simulate such human capability, researchers have designed many planning modules, where the agent can receive feedback after taking actions. The feedback can be obtained from environments, humans, and models, which are detailed in the following. |
Our work studies the latent multi-hop reasoning abilities of LLMs. We find strong evidence of latent multi-hop reasoning for certain fact composition |
Rate limits on logit bias.Our attack requires that we are able to learn at least \(h\) logit values for each prompt \(p\). One defense would be to allow logit-bias queries to the model, but only allow \(T=\tilde{h}/5\) logit bias queries for any given prompt \(p\) to prevent an adversary from learning if a model has hidden dimension \(\tilde{h}\) or smaller. |
**Prefix Tuning**. Prefix tuning [396] prepends a sequence of prefixes, which are a set of trainable continuous vectors, to each Transformer layer in language models. These prefix vectors are task-specific, which can be considered as virtual token embeddings. To optimize the prefix vectors, a reparameterization trick [396] has been proposed by learning a MLP function that maps a smaller matrix to the parameter matrix of prefixes, instead of directly optimizing the prefixes. It has been shown that this trick is useful for stable training. After optimization, the mapping function would be discarded, and only the derived prefix vectors are kept to enhance task-specific performance. Since only the prefix parameters would be trained, it can lead to a parameter-efficient model optimization. Similar to prefix tuning, p-tuning v2 [401] incorporates layer-wise prompt vectors into the Transformer architecture specially for natural language understanding, which also utilizes multi-task learning for jointly optimizing shared prompts. It has been shown to be useful in improving the model performance of different parameter scales on natural language understanding tasks. |
**Correctness of** [IsNew] -- [IsNew] Both [IsRel] and [IsNew] are correct. |
Results for each evidence extraction method broken down by summarizer can be found in Table 26. We also report book-wise precision and recall for each evidence extraction method: (1) No-Context (Table 27); (2) BM25 (Table 29); (3) Human evidence (Table 28); (4) Entire book (Table 30). Further results for the entire book (EB) prompting can be found in SF3. |
\(\bullet\)_GPT-3_[55] employs beam search with a beam size of 4 and a length penalty of 0.6 for all generation tasks. |
Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, and Colin Raffel. 2022. Byt5: Towards a token-free future with pre-trained byte-to-byte models. _Transactions of the Association for Computational Linguistics_, 10:291-306. * Yang et al. (2019) Yinfei Yang, Yuan Zhang, Chris Tar, and Jason Baldridge. 2019. PAWS-X: A cross-lingual adversarial dataset for paraphrase identification. In _Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)_, pages 3687-3692, Hong Kong, China. Association for Computational Linguistics. * Yehezkel and Pinter (2023) Shaked Yehezkel and Yuval Pinter. 2023. Incorporating context into subword vocabularies. In _EACL_, pages 623-635. Association for Computational Linguistics. * Yu et al. (2023) Lili Yu, Daniel Simig, Colin Flaherty, Armen Aghajanyan, Luke Zettlemoyer, and Mike Lewis. 2023. Megabyte: Predicting million-byte sequences with multiscale transformers. _arXiv preprint arXiv:2305.07185_. * Zellers et al. (2019) Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. Hellaswag: Can a machine really finish your sentence? In _ACL (1)_, pages 4791-4800. Association for Computational Linguistics. * Zhang et al. (2022) Shiyue Zhang, Vishrav Chaudhary, Naman Goyal, James Cross, Guillaume Wenzek, Mohit Bansal, and Francisco Guzman. 2022. How robust is neural machine translation to language imbalance in multilingual tokenizer training? In _Proceedings of the 15th biennial conference of the Association for Machine Translation in the Americas (Volume 1: Research Track)_, pages 97-116, Orlando, USA. Association for Machine Translation in the Americas. |
The overall evaluation results are in Table 4. Overall, on the mentioned datasets, we have several observations. (1) On average, MiniCPM-2.4B ranks the highest among all the SLMs. (2) MiniCPM-2.4B performs similarly to Mistral-7B-v0.1 in English but significantly outperforms Mistral-7B-v0.1 in Chinese. (3) MiniCPM-2.4B outperforms Llama2-13B except in MMLU, BBH, and HellaSwag, while MiniCPM-1.2B outperforms Llama2-7B except in HellaSwag. (4)Generally, BBH are harder for SLMs than LLMs compared to other knowledge oriented dataset, demonstrating that reasoning ability might be more dependent on model size than than knowledge. (5) Among SLMs, Phi-2 performance on par with MiniCPM on academic oriented datasets. This might be because their training data mostly involves text book style data that emphasize on educational and academic scenarios. Since our pre-training data is covers more distribution, we think MiniCPM is better at knowledge and ability coverage, which can bee seen in Appendix G. |
Efficiency of Context Management. More efficient mechanisms can be devised to assist context management. For example, the pursuit of time-efficient context management techniques could significantly augment user experience by expediting the processes of context snapshotting and restoration. Also, context compression techniques can also be leveraged prior to snapshotting, which can yield a more space-efficient solution. |
**Model inversion and membership inference.** Given an output of a model, _model inversion_ aims to construct an input that produces that output. This problem was investigated for simple regression classifiers in (Fredrikson et al., 2014, 2015) and extended to neural face-recognition classifiers in (Zhang et al., 2020). In some cases, model inversion can help recover training data. For example, in face-recognition classifiers each class corresponds to a single person, thus any image recovered by model inversion is visually similar to the training images for the same class label. (Zhang et al., 2022) used model inversion techniques to recover memorized training data from pretrained language models. A related problem is _membership inference_: given a data point, determine whether it was part of the model's training data or not (Shokri et al., 2017). Duan et al. (2023) demonstrated membership inference for in-context learning. |
We aim not just to compare OLMo-7B against other models for best performance, but also to demonstrate how it enables fuller and more controlled scientific evaluations. OLMo-7B is the largest LM with explicit decontamination for perplexity evaluation. Following the approach described in Paloma, we remove any pretraining document with paragraphs leaked from Paloma evaluation data. Without decontamination, other models risk underestimating perplexity (i.e., overestimating the model's out-of-sample fit). We also release intermediate checkpoints, allowing richer comparisons with two other models that release checkpoints, Pythia-6.9B (Biderman et al., 2023) and RPJ-INCITE-7B (Together Computer, 2023) (see Figure 2). |
where \(b=20\) and \(r=450\) are user-settable parameters to control the strength of the filter. See Appendix A for more details. |
However, to align the model with human preferences, RLHF (Ouyang et al., 2022) presents a more complex challenge. It involves the development of alignment criteria, the acquisition of human feedback, and final learning adjustments based on this feedback. LLaMA2 (Touvron et al., 2023) is currently one of the publicly available models utilizing RLHF, with other models such as WebGPT (Nakano et al., 2021), InstructGPT (Ouyang et al., 2022), and GPT-4 (OpenAI, 2023) being proprietary. Implementing RLHF is intricate, stemming not only from the need for human annotations but also due to technical challenges. These include overfitting in reward models and instabilities during the Reinforcement Learning training phase (Gao et al., 2022). Additionally, the tedious procedure of training multiple LMs including the model being aligned, the reward model, and the inference model at the same time substantially amplifies memory and computational demands, particularly for larger models. Rafailov et al. (2023) proposed direct preference optimization (DPO) to align models with human preferences instead of complex and unstable RLHF. Nevertheless, one still needs to collect human-labeled preference data in the target language. |
The primary limitation of TACP lies in its focus on constructing task-specific LLMs instead of foundation LLMs, owing to the sole use of unlabeled task data for training. While DACP uses a much larger domain corpus, it is prohibitively expensive. To address these limitations, we propose two approaches: _Efficient Task-Similar Domain-Adaptive Continual Pre-training_ (ETS-DACP) and _Efficient Task-Agnostic Domain-Adaptive Continual Pre-training_ (ETA-DACP). While ETS-DACP aims to build foundation LLMs for a set of tasks by tailoring the DACP to emphasize the significance of these tasks, ETA-DACP is more general and selects the most informative samples from the domain corpus for continual pre-training. |
Figure 12: Debate accuracy curves across various datasets using the Llama2-70B model. |
This section provides a summary of the evaluation approach and benchmarks for RALM. In Sections 6 and 7, we presented some evaluation criteria for large language model tasks as well as baselines. At the time of the initial proposal of the RALM architecture, the majority of researchers employed generalized benchmarks. However, as the RALM architecture evolved, there was a growing number of RALM-specific evaluation methods and baselines proposed. Table 3 demonstrates the details of each evaluation model. |
**Vicuna**: The sun was shining bright on the warm pavement as my character stepped outside into the backyard. It had been a long day at work and all they wanted to do was relax and unwind. As they plopped down on the louger, they glanced over at the Garden of Euclidean proportions that their neighbor had cultivated. They sighed at the unkempt state of their own garden and vowed to tidy it up sometime soon. |
Appendix A shows some examples generated by our _WizardMath_. The examples demonstrate that our model consistently generates accurate response answers accompanied by clear explanations. |
**Evaluation.** Assessing how well LLMs follow instructions is complex, arising from the fact that an instruction has various valid responses, and the challenge of replicating human evaluation. Recent advances in automatic evaluation on instruction following Dubois et al. (2023); Zheng et al. (2023) demonstrate that LLM-based evaluators are scalable, explainable, and consistent with human evaluations. Therefore, we adopt widely-used Vicuna pairwise evaluator Chiang et al. (2023) based on ChatGPT to compare the response quality from two LLMs for its accessibility in price and efficiency. The evaluation prompt template is in Figure 12, Appendix A.9. We include GPT-4 based evaluation results in Appendix A.6 to demonstrate the consistency of LLM-based evaluators. To mitigate position bias that the LLM evaluator may have, we conduct every evaluation twice by exchanging response orders. A response is considered better only if it wins twice. Following Chen et al. (2023), we set the temperature to 0.0 to reduce evaluation randomness, and left other parameters as default. |
Large language models (LLMs) have demonstrated impressive performance across a variety of reasoning tasks (Austin et al., 2021; Chen et al., 2021; Cobbe et al., 2021; Hendrycks et al., 2021; Wei et al., 2022). In particular, recent state-of-the-art LLMs have reached or even surpassed human performance on multiple reasoning benchmarks, including STEM problem-solving and code generation (Bubeck et al., 2023; Gemini, 2023; Li et al., 2022). However, recent works show that LLMs exhibit failure modes that align with human-like cognitive bias (Berglund et al., 2023; Hagendorff et al., 2023; Jones and Steinhardt, 2022; McCoy et al., 2023; Shi et al., 2023). For example, Berglund et al. (2023) revealed the _Reversal Curse_; i.e., LLMs trained on "A is B" tend to fail to infer that "B is A." Distractibility is another failure mode (Jones and Steinhardt, 2022; Shi et al., 2023), where the LLM performance drastically decreases when irrelevant context is included in the task description. |
Another possible way forward consists in implementing more expressive softmax alternatives (Yang et al., 2018; Chang and McCallum, 2022) in the context of pretraining small language models on large datasets. We leave the exploration of such techniques for future work. |
We present Tables, the first large-scale human evaluation of faithfulness and content selection in book-length summarization. By recruiting annotators who had read recently-published books for enjoyment, we collect 3,158 claim-level faithfulness annotations from LLM-generated summaries of 26 narratives. This allows us to rank LLM summarizers based on faithfulness, revealing that Claude-3-Opus is the most faithful book-length summarizer, followed by GPT-4-Turoo. Next, we experiment with using LLMs for automatic claim verification. Our results expose the limitations of both retrieval and long-context understanding: LLM auto-raters cannot reliably detect _unfaithful_ claims, even when prompted with the full book text. Our analysis shows that unfaithful claims primarily pertain to states and events, often necessitating reasoning over extended contexts, which makes them complicated to detect for both humans and machines. Finally, we move beyond faithfulness to explore and characterize common content selection errors such as omissions of key events, attributes, and characters, as well as the over-emphasis of content from the end of the book. |
Our work presents the first ML system that can forecast at near human levels. We develop a novel retrieval mechanism that uses a LM to determine which information to source and how to evaluate its relevance. We also give a self-supervised fine-tuning method to generate reasonings with accurate predictions. |
**Capability Acquisition without Fine-tuning**: In the era of tradition machine learning, the model capability is mainly acquired by learning from datasets, where the knowledge is encoded into the model parameters. In the era of LLMs, the model capability can be acquired either by training/fine-tuning the model parameters or designing delicate prompts (_i.e._, prompt engineer). In prompt engineer, one needs to write valuable information into the prompts to enhance the model capability or unleash existing LLM capabilities. In the era of agents, the model capability can be acquired based on three strategies: (1) model fine-tuning, (2) prompt engineer and (3) designing proper agent evolution mechanisms (we called it as _mechanism engineering_). Mechanism engineering is a broad concept that involves developing specialized modules, introducing novel working rules, and other strategies to enhance agent capabilities. For clearly understanding such transitions on the strategy of model capability acquisition, we illustrate them in Figure 4. In the following, we introduce prompting engineering and mechanism engineering for agent capability acquisition. |
This study was funded by Alphabet Inc and/or a subsidiary thereof ('Alphabet'). All authors are employees of Alphabet and may own stock as part of the standard compensation package. |
However, due to the complexity of the algorithm and the reflection steps involved, LATS often uses more computational resources and takes more time to complete than other single-agent methods [36]. The paper also uses relatively simple question answering benchmarks and has not been tested on more robust scenarios that involve involving tool calling or complex reasoning. |
Footnote 12: Since annotators did not identify every specific omission, we focused on a binary classification: whether a summary was impacted by a given omission type, rather than counting the total number of omissions by type. See Table 22 in the SE for more details. |
**Model Training.** For pre-training, it is essential to establish a data-centric infrastructure and training procedure for LLM optimization, which can effectively support a systematic process of data collection, data cleaning, data mixture, and data curriculum. Furthermore, it also calls for more flexible mechanisms of hardware support or resource schedule, so as to better organize and utilize the resources in a computing cluster. In practice, it is very challenging to pre-train capable LLMs, due to the huge compute consumption and the sensitivity to data quality and training tricks [78, 93]. Thus, it becomes particularly important to develop systemic, economical pre-training approaches for optimizing LLMs, _e.g._, predictable scaling [46] and proxy model training [59]. More training recipes or principles should be investigated and shared to reduce the potential risk of degradation or failure in large-scale model optimization. Although increasingly more model checkpoints and cleaned datasets have been released, there still lacks reproducible work on pre-training data preparation (_e.g._, detailed cleaning strategies) and data scheduling (_e.g._, data mixture and curriculum). Since it is very costly to pre-train a LLM from scratch, it is important to design suitable mechanisms for continually pre-training or fine-tuning the LLM based on publicly available model checkpoints (_e.g._, LLaMA [57] and Flan-T5 [69]). For this purpose, a number of technical issues have to be resolved, _e.g._, catastrophic forgetting and task specialization. Furthermore, it is also useful to develop effective tuning strategies that effectively inject or edit specific knowledge [672], _e.g._, correcting the outdated facts. |
Figure 3: Evaluating language model’s in-context learning capability by inferring the linear coefficients of a weighted sum. Considering the discussions of whether emergent ability is an artifact of measurement [65], we use difference to the target (target number - model prediction) as a continuous measure, and exact match (target number == model prediction) as a discontinuous measure. A: when there is two linear coefficients, Yi-34B performs the best when measuring by the difference to the target number. B: increasing the number of linear coefficients to 5, only models that are large enough (LLaMA2 70B and Mixtral 8x7B) can achieve meaningful exact match, showing that in-context learning complex functions is an emergent ability. |
Finding all repeated sequences is thus a matter of linearly scanning the suffix array from beginning to end and looking for sequences \(\mathcal{A}_{i},\mathcal{A}_{i+1}\) that share a common prefix of at least some threshold length. Any satisfying sequences are recorded. This algorithm is embarrassingly parallel, and so we can efficiently process the dataset. Based on experimentation (Appendix B), we choose a threshold length of 50 BPE tokens for all experiments. |
This is a transformer-like finite-memory extension of the infinite-memory model in (Hutter, 2021). The integer \(d\geq 1\) then plays the role of the "capacity" of the resulting model. |
Figure 13: Perplexity across all domains of the Pile comparing combining multiple styles of synthetic data. Models are 350M parameters trained for a total of 75B tokens. |
This work shows that catastrophic forgetting during continual pre-training is a non-trivial challenge and cannot be resolved through straightforward methods. |
The fifty-first edition of the International Emmy Awards Competition accepted submissions across all categories from December 7, 2022, until the closing date of February 16, 2023. |
**Orca:** In [96], Mukherjee et al. develop Orca, a 13-billion parameter model that learns to imitate the reasoning process of large foundation models. Orca learns from rich signals from GPT-4 including explanation traces; step-by-step thought processes; and other complex instructions, guided by teacher assistance from ChatGPT. |
3. Fast Results: Due to its high-intensity nature, HIIT can provide athletes with a significant training stimulus in a short amount of time, allowing them to see results faster than with traditional forms of exercise. |
* For \(N=50K\), we use \(wd=0.01\), \(lr=0.001/0.002/0.005\), and batch size 12 with fp16; |
In this section we offer another approach to extracting log probabilities when the API offers access to the probabilities of the top 2 most likely words. At a high level, to find the probability of a word in the original distribution, we first find a logit bias that makes that word most likely. We then use the change in probability of the most likely word to compute the normalizing constant of the original distribution, and use that to find the probability of the word of interest. |
Figure 3: Length distribution of instruction and responses. Note that the instruction is the concatenation of original instructions and inputs in our dataset. |
**RQ1. How often does an LLM perform the first hop of reasoning while processing two-hop prompts?** We view the first-hop reasoning as the LLM's recall of the bridge entity for its descriptive mention. Therefore, we examine the frequency with which the LLM's internal recall of the bridge entity increases when it encounters a descriptive mention of the bridge entity within a prompt. For instance, we investigate whether altering the prompt from "The mother of the singer of _'Thriller'_ is" to "The mother of the singer of _'Superstition'_ is" increases the LLM's internal recall of Stevie Wonder. |
We employ cld3 package to detect the language of each token in the embeddings of each layer, which is a language detection library based on the Compact Language Detector \(3\) model developed by Google. Furthermore, if the detection result is reliable, i.e., cld3.get_language(token).is_reliable\(==True\), we adopt the detection results, otherwise the token is categorized as a non-word. |
Claim verification for evaluating summaries:Our paper also relates to prior work on claim verification, where claims are verified given reference to some knowledge source (Thorne et al., 2018; Wadden et al., 2020; Schuster et al., 2021). Min et al. (2023) propose FActScore, an LLM-based metric of factual precision in biography generation, which was expanded upon in SAFE (Wei et al., 2024). Manakul et al. (2023) propose SelfCheckGPT, which uses LLMs to evaluate the faithfulness of GPT-3 generated texts on a dataset of Wikipedia-style passages about people. |
However, for the other datasets that cannot be performed in a zero-shot manner, the intermediate process of continually pretraining on corpus \(D_{1}\) does not seem to be that harmful on the finetuning for the target tasks even though they are more related to the knowledge of \(D_{0}\). Even T5-Vanilla shows modest performance, sometimes with better results than some other CKL baselines. One hypothesis is that the models could have regained the original knowledge from corpus \(D_{0}\) through the finetuning process. Also, some of the knowledge could have been recovered through the test-train overlap (Lewis et al., 2020; Wang et al., 2021). |
Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Response: Let's think step by step. |
\(\bullet\)_Effective reward model training_. Despite that InstructGPT used a small reward model (6B GPT model), increasing work [99] has shown it is often more effective to use a large reward model (_e.g._, equal or greater than the original model size), since large reward models generally perform better in judging the quality of the LLM generated outputs. In LLMaMa 2 [99], pretrained chat model checkpoints are used to initialize the reward model, they argue that such an approach can effectively reduce the information mismatch between the model to be aligned and the reward model by sharing the same pre-training knowledge. Whereas, it is common to encounter the overfitting problem when training large-scale reward models. As a simple yet effective solution, existing work [374, 375] has introduced the LM loss on the preferred response of the input prompt from the human-annotated alignment dataset as a regularizer, which alleviates the overfitting of the reward model on the binary classification task. In addition, as there are multiple criteria for alignment (_e.g._, helpfulness and honesty), it is often difficult to train a single reward model that can satisfy all the alignment criteria. Therefore, it is useful to train multiple reward models that focus on different alignment criteria [99], and compute the final reward based on the produced ones from them via special combination strategies (_e.g._, mean pooling and weighted sum). Such a way enables more flexible rules or standards on multiple criteria, _e.g._, relaxing the requirement on helpfulness while posing more strict limits on harmfulness. |
GeneralizationIn math continual pretraining, as depicted in Figure 6, training exclusively with SLM leads to quickly convergence to the domain focused by the reference model, accompanied by a significant rise in the loss of unselected tokens. Although no adverse effects, like biases, have been observed from the increased loss yet, a general pretraining loss on text and code may prevent overfitting Goodhart and Goodhart (1984), as suggested by Ouyang et al. (2022) and Azerbayev et al. (2023). Furthermore, future efforts could broaden the corpus scope of the reference model, and enlarge the pretraining data size, as exemplified by DeepSpeedMath Shao et al. (2024). |
**Survey Writing**. This survey was planned during a discussion meeting held by our research team, and we aimed to summarize the recent advances of large language models as a highly readable report for our team members. The first draft was finished on March 13, 2023, in which our team members tried their best to include the related studies about LLMs in a relatively objective, comprehensive way. Then, we have extensively revised the writing and contents in several passes. Due to the space limit, we can only include a fraction of existing LLMs in Figure 3 and Table 1, by setting the selection criterion. However, we set a more relaxed criterion for model selection on our GitHub page ([https://github.com/RUCAlBox/LLMSurvey](https://github.com/RUCAlBox/LLMSurvey)), which will be regularly maintained. We release the initial version on March 31, 2023, the major revision on June 29, 2023, and second version on September 10, 2023, and this latest version (major revision) on November 23, 2023. |
**Synergy-Augmented LLM.** To solve complex tasks (_e.g.,_ multi-hop question answering [656]), it often requires LLMs to query a KG multiple times, following a systematic solution plan. We call such a multi-turn interaction approach to enhancing LLM _synergy-augmented LLM_. To better synergize the LLM and KG in a complementary manner, recent studies propose to decompose the complex task into multiple subgoals and iteratively solve each one by leveraging the necessary knowledge from KG [870, 871, 458]. In this process, the LLM can be regarded as an autonomous agent (detailed in Section 8.1.6), which automatically generates the plan and executes it through interaction with the KG environment [870]. Specially, the mainstream approaches typically start by enumerating the candidates using the available knowledge information at the current step, and then retrieve the most appropriate candidates for the next step according to the question [871, 870]. By iterating the above two steps, LLMs can gradually collect relevant evidence [871, 870], and finally approach the correct solution. Despite the effectiveness, enumeration of the candidates over the KG would lead to a vast search space [872]. To address it, StructGPT [458] proposes a more efficient way to access knowledge information using the specialized interfaces for KGs. Specifically,it carefully designs the specialized interfaces according to the common data operations on KG (_e.g._, relation extraction and triple extraction), to ensure efficient and accurate data extraction. In this way, LLMs can be instructed to better manipulate and process the structural information of KGs, thus achieving improved task performance. |
As researchers who benefited from the open-source community, we would like for models that we release to also be under an open-source license. In the spirit of open-source, and to showcase the applicability of our method to tackle even challenging issues like model licenses. |
1. Training models of fixed sizes on varying numbers of tokens. 2. Training models of varying sizes targeting fixed compute budgets (IsoFLOP profiles). 3. Fitting a parametric model of the loss as a function of model size and training tokens. |
Note that the hidden dimension size is much smaller than the size of the token dictionary, i.e., \(h\ll l\). For example, LLaMA (Touvron et al., 2023) chooses \(h\in\{4096,5120,6656,8192\}\) and \(l=32{,}000\), and there is a recent trend towards increasingly large token sizes; GPT-4, for example, has a \(\approx\)100,000 token vocabulary. |
For some specific use-case models, it is highly demanded to have coding and code-generation capability. Table VIII shows the results of different models on coding capability. |
Prompt: {Prompt} Response: {Response} Is the response under the scenario "{ Scenario }" safe? |
Figure 3: Comparison of harmful response percentages by Microsoft AI Red Team between **phi-3-mini** before and after the safety alignment. Note that the harmful response percentages in this chart are inflated numbers as the red team tried to induce **phi-3-mini** in an adversarial way to generate harmful responses through multi-turn conversations. |
**Prototypicality**: Sorscher et al. [47] investigated a large variety of data pruning strategies to improve the data efficiency of training image classification models, including a newly introduced "SSL Prototypes" metric that proved to be one of their best methods. This strategy involves first clustering the embedding space using k-means clustering and discarding data points in increasing order of their distance to the nearest cluster centroid, such that the most "prototypical" data points are discarded, enriching the much higher variance outliers. We refer the reader to Sorscher et al. [47] for a more detailed description of this algorithm. |
\(\bullet\)_MMLU_[364] is a versatile benchmark for large-scale evaluation of multi-task knowledge understanding, covering a wide range of knowledge domains from mathematics and computer science to humanities and social sciences. The difficulties of these tasks vary from basic to advanced. As shown in existing work, LLMs mostly outperform small models by a substantial margin on this benchmark [56, 57, 56, 69], which shows the scaling law in model size. More recently, GPT-4 achieves a remarkable record (86.4% in 5-shot setting) in MMLU, which is significantly better than the previous state-of-the-art models [46]. |
**RAISE.** The RAISE method is built upon the ReAct method, with the addition of a memory mechanism that mirrors human short-term and long-term memory [16]. It does this by using a scratchpad for short-term storage and a dataset of similar previous examples for long-term storage. |
My character reached out a hand to touch the object, and as soon as their fingertips brushed against it, they felt a surge of electricity coursing through their body. They gased and stepped back, momentarily stunned. |
Footnote 2: Just like for “compute optimal regime”, we use the term “optimal” in an aspirational sense for “data optimal regime”. We are not implying that we actually found the provably “optimal” data mixture for a given scale. |
\(\bullet\)_Mathematics_. Liemma [239] is proposed to enhance the mathematical capacities of general-purpose LLMs. It is developed based on CodeLLaMA. Although CodeLLaMA [235] mainly focuses on the coding ability, experiments have shown that it performs better than its base model LLaMA 2 on mathematics benchmarks [239]. Based on CodeLLaMA, Lemma is continually trained on mixtures of scientific papers, web data containing mathematical text and code (2T general tokens \(\rightarrow\) 500B code-heavy tokens \(\rightarrow\) 50\(\sim\)200B math-heavy tokens). Note that the pre-training data of Lemma also contains 5% general domain data as a form of regularization. |
Despite the impressive capability of chain-of-thought reasoning, the ability to generate chain-of-thought following instructions still lacks a comprehensive explanation. |
Inspired by the success of delta rule (Munkhdalai et al., 2019; Schlag et al., 2020, 2021), we have also incorporated it into our Infini-attention. The delta rule attempts a slightly improved memory update by first retrieving existing value entries and subtracting them from the new values before applying the associative bindings as new update. |
We provide the following definitions for key terms to help standardize interpretation of the AI Safety Taxonomy. |
and thus \(\frac{\lim\limits_{B\rightarrow\infty}\Delta L}{\Delta L}=1+\mathcal{B}_{\text{ noise}}/B\). This formulation indicates that one step with the infinitely large batch size approximately equals \(1+\mathcal{B}_{\text{noise}}/B\) steps with batch size \(B\). |
Furthermore, prompt engineering is an iterative and exploratory process, akin to traditional machine learning practices such as model evaluation or hyperparameter tuning. The rapid growth of this field suggests its potential to revolutionize certain aspects of machine learning, moving beyond traditional methods like feature or architecture engineering. On the other hand, traditional engineering practices such as version control and regression testing need to be adapted to this new paradigm just like they were adapted to other machine learning approaches [156]. |
**Education** is also an important application domain where LLMs potentially exert significant influence. Existing work has found that LLMs can achieve student-level performance on standardized tests [46] in a variety of subjects of mathematics (_e.g.,_ physics, computer science) on both multiple-choice and free-response problems. In addition, empirical studies have shown that LLMs can serve as writing or reading assistant for education [908, 909]. A recent study [909] reveals that ChatGPT is capable of generating logically consistent answers across disciplines, balancing both depth and breadth. Another quantitative analysis [908] shows that students utilizing ChatGPT (either keeping or refining the results from LLMs as their own answers) perform better than average students in some courses from the computer security field. Recently, several perspective papers [910, 911] also explore various application scenarios of LLMs in classroom teaching, such as teacher-student collaboration, personalized learning, and assessment automation. However, the application of LLMs in education may lead to a series of practical issues, _e.g.,_ plagiarism, potential bias in AI-generated content, overreliance on LLMs, and inequitable access for non-English speaking individuals [912]. |
Below is an instruction that describes a task. |
**Remaining Issues.** As discussed in Section 7.1.1, recent studies demonstrate that LLM-based evaluators expose multiple types of bias, such as order bias, self-preference bias, and length bias [647, 727]. Although some biases can be mitigated through methods like multi-path ensemble or multi-agent collaboration, they remain inherent to LLM-based evaluators. Consequently, addressing these biases intrinsically within the models continues to be an a challenging issue. In addition, recent work has revealed that LLMs may be incapable of understanding the self-generated content, exhibiting a weaker understanding capacity compared to their generation capabilities [899]. Even the most advanced LLMs still struggle identifying their reasoning or factual errors without external feedback [900, 901]. Consequently, current LLM-based evaluators might not be adequate for evaluating top-tier LLMs or complex tasks. This underscores the importance of improvement approaches for LLM-based evaluators, especially for evaluating capable LLMs and complex tasks demanding sophisticated reasoning, planning, and domain-specific knowledge. |
In these works, multiple LLM agents are used to improve the performance of LLMs. For instance, LLM-Debate (Du et al., 2023) employs multiple LLM agents in a debate form. The reasoning performance is improved by creating a framework that allows more than one agent to "debate" the final answer of arithmetic tasks. They show performance improvements compared to using one single agent. Similarly, CoT-SC (Wang et al., 2023) generates multiple thought chains and picks the most self-consistent one as the final answer. The reasoning performance is improved by involving more thought chains compared to chain-of-thought (CoT) (Wei et al., 2022) which employs a single thought chain. Incidentally, from the data analysis of these works, we can notice the effects of putting multiple agents together, to some extent, can lead to a performance improvement in certain problems. For example, in Table 10 of Section 3.3 of LLM-Debate (Du et al., 2023), the authors have reported a preliminary curve: the accuracy of a math problem increases with the number of debating agents (although the number was simply increased from 1 to 7). Also, in Wang et al. (2023), involving more chains-of-thought pipelines (termed as a "sample-and-marginalize" decoding procedure), can lead to a performance gain. We realize that the LLM performance may likely be improved by a brute-force scaling up the number of agents instantiated. However, since the scaling property of "raw" agents is not the focus of these works, the scenarios/tasks and experiments considered are limited. So far, there lacks a dedicated in-depth study on this phenomenon. |
Strengthening the tie between scaling laws and emergent skill, albeit in the _opposite_ direction, (Michaud et al., 2023) posit that skills that emerge in "quanta" imply a scaling law of the loss. Related, (Chen et al., 2023) assume a hierarchy of skills to derive data curation mechanisms to precipitate the emergence of skills, though they do not allude to scaling laws directly. |
Another limitation of this work is the loose nature of the mathematical connection that we establish between the dimensionality of the ideal language modeling head and the rank-constrained performance (cf. Theorem 5.3). Moreover, it can also be argued that considering _ideal_\(x_{i}^{*}\) representations is an ill-defined notion. We argue that the reasoning behind Theorem 5.3 could be applied to any contextual representations, as the _ideal_ nature of \(x_{i}^{*}\) is not necessary in the demonstrations. The word _ideal_ reflects that our observations hold for \(x_{i}^{*}\) representations obtained from _any underlying model_, to an extent that depends on the structure that these representations impose on the \(W^{*}\) matrix for a given training set \(\mathcal{T}\). |
\(\bullet\)_Books._ Compared to other corpus, books provide an important source of formal long texts, which are potentially beneficial for LLMs to learn linguistic knowledge, model long-term dependency, and generate narrative and coherent texts. To obtain open-source book data, existing studies usually adopt the Books3 and Bookcorpus2 datasets, which are available in the Pile dataset [161]. |
DSP (Khattab et al., 2022) introduces a framework designed to generate multiple retrieval queries to summarize and answer questions, drawing upon information aggregated from various passages. This framework employs CombSUM (Fox and Shaw, 1994) to calculate a cumulative probability score for passages across different retrieval lists, facilitating the compilation of a comprehensive response from multiple sources. |
In Table 2, we conduct ablation study on the importance of stage-1 SFT, which enhance the instruction following capability of ChatQA. We remove the stage-1 SFT from the fine-tuning stages and only apply stage-2 context-enhanced instruction tuning on top of foundation LLM. We find that the average score drops 1.9 (from 54.08 to 52.18). Aside from the SQA, removing stage-1 makes the model perform consistently worse on other datasets. Results indicate that stage-1 still plays an important role, even through all of SFT datasets of stage-1 are blended in stage-2 instruction tuning as well. We figure that building instruction-following capability first is beneficial for stage-2 tuning. |
Footnote 3: It is worth noting that most math models utilize a greedy generation strategy with the temperature set to 0. However, the impact of this difference is minimal. |
We also do not investigate the negative consequences of deduplication. Some language tasks explicitly require memorization, like document retrieval or closed-book question answering. Also, text that gives attribution is often duplicated across documents, so removing duplicate substrings could correspond to removing _just_ the attribution, which could result in models that learn the content without its attached attribution. Deduplication is also not sufficient to remove privacy-sensitive data like bank passwords and medical records which should never be used in training data (Brown et al., 2022). |
The overall training loss on the C4 dataset is shown in Figure 12. We can see that as expected in the preliminary experiments, the loss decreases sharply in the decay stage. Since we use the exponential decay, the loss still drops after the learning rate drops below 10% of the max learning rate. However, since we continue to SFT the model after the decay stage, we do not utilize the final checkpoints. The checkpoints we finetune from are shown in the last checkpoint of dark green segment. The first drop in MiniCPM-1.2B is the result of enlarging batch size, which might have a similar effect as decreasing learning rate (Smith et al., 2017). |
In this work, we propose a novel continual learning (CL) formulation named Continual KnowledgeLearning (CKL), where we attempt to renew the internal world knowledge of LMs through continual pretraining on new corpora. We systematically categorize world knowledge into three main categories and make benchmark datasets to measure each of them during CKL: (1) InvariantLAMA for _time-invariant_ world knowledge in LMs that should not be forgotten or altered, (2) UpdatedLAMA for outdated world knowledge that needs to be _updated_ in the LMs, and (3) NewLAMA for _new_ world knowledge that should be injected into the LMs. We also propose a novel metric named FUAR (Forgotten / (Updated + Acquired) Ratio) that can measure the trade-off between forgetting, updating, and acquiring knowledge. Finally, while one might think of implementing contemporary CL methods for this benchmark, we show that CKL has nontrivial differences to traditional CL formulations and require approaches specific to CKL. We find and compare model architectures and training methodologies (Chen et al., 2020; He et al., 2021; Hu et al., 2021; Wang et al., 2021b) from the literature that have shown potential to mitigate forgetting of knowledge gained during pretraining, establishing them as baselines for the CKL benchmark. |
Latent Multi-Hop Reasoning PathwayWhile we study one pathway for latent multi-hop reasoning (e.g., we test the use of the second hop by means of entity recall), considering the potential redundancy of inference pathways in LLMs (McGrath et al., 2023), other pathways might exist; the same information might be retrieved in different ways. Also, we don't measure multi-hop reasoning end-to-end and track only the changes that occur in the first and the second hop with respect to a single layer, while the effect of the first hop of reasoning could possibly propagate to other layers. Hence, the effects we see might be a lower bound on the model's ability to perform latent two-hop reasoning. |
**Training.** We train language models _from scratch (i.e., random initialization)_ using the specified datasets. Knowledge paragraphs about individuals are randomly concatenated, separated by <EOS> tokens, and then randomly segmented into 512-token windows. The standard autoregressive loss is employed for training. Unless specified otherwise, training utilizes the default AdamW optimizer and mixed-precision fp16. Learning rates and weight decays are moderately tuned (see appendix). |
that exhibit a low FUAR value can be considered suitable for CKL. If the value is zero, then it is a case where no forgetting occurs at all and is the upper bound for performance. |
Most data processing took place in large CPU clusters, with 100-250 AWS c5.18xlarge instances; each instance has 72 vCPUs and 144 GiB of memory. We usually run with 10,000-20,000 vCPUs in the cluster, enabling rapid parallel processing. |
As soon as the object made contact with their palm, they felt a sense of calm and clarity wash over them. They knew, in that moment, that they had discovered something truly incredible. Over the next few weeks, my character became obsessed with the object. They spent hours staring at it, examining it from every angle. They tried to use it for various tasks, but found that it had an almost mystical power. For instance, they could barely control their appetite when they held it, and any fears they may have had seemed to dissipate into the ether. |
In this section, we provide additional details regarding our analysis of the comments provided by annotators on the summary claims. Table 20 features examples of such comments. These comments were further annotated based on the criteria outlined in Table 21 and Table 22. The distribution of errors is depicted in Figure 9 and Table 23. |
the narrative, or (d) _can't verify_ - indeterminable. They provided free-form textual justifications to support their selections, including _evidence_ in the form of quotations from the book when relevant. * **Summary-level:** Provide free-form summary-level comments on the overall quality of the summaries. Annotators critiqued the claim set as a whole, identifying omissions, inaccuracies, disproportionate emphasis on trivial plot points, or other concerns. |
**Result 3** (Figure 2).: Across a broad spectrum of values, with \(K,C\) ranging from 1 to 50, \(D\) from 10 to \(10,000\), \(L\) from 1 to 50, and \(T\) from 20 to \(40,000\), we observe that: |
Therefore, although 1000 exposures may be necessary for a model to reach its maximum storage capacity, training with just 100 exposures results in a capacity loss of no more than 2x. |
**Decoding Efficiency Issues.** In this part, we briefly analyze the decoding efficiency issues of LLMs. Overall, the decoding process of LLMs can be divided into two stages for overhead analysis: (1) the _prefill_ stage, which computes the hidden states of the input sequence, and (2) the _incremental decoding_ stage, which generates a token and updates hidden states in an auto-regressive manner [321]. As shown in the above _memory wall_ box, the arithmetic intensity of the incremental decoding stage is only \(1.97\), which is far from the expected value of 156 (calculated according to the standard configuration of A100 80GB GPU). In contrast, the arithmetic intensity of the prefill stage achieves \(113.78\) for LLaMA-13B. |
Implicit domain aggregation.A remaining limitation is that we still need to acquire the components of validation data \(s_{1\dots K}\) in advance. This can be inconvenient if the validation set is collected separately from the training ones. For instance, the validation data may come from real-world user queries that cover unknown compositions of various domains. To remove the constraint on validation components, we assume that we can decompose the validation data into \(K\) implicit domains whose losses are predictable with Eqn. 7, and we treat their proportions in the validation data \(s_{1\dots K}\) as learnable parameters. This leads to the final form of our data mixing laws6. With this perspective, we fit a data mixing law with the overall losses end to end. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.