Datasets:
Translation-EnKo
/
arxiv-translation

Dataset card Data Studio Files Community
2
text
stringlengths
4
4.47k
Planning augmented approaches have advanced conventional sequential planning by introducing search-based, graph-based, and definition language-based methods. On the other hand, some methods incorporate action, planning, reflection, or tools, aiming to enhance LLMs' long-term planning and error resilience capabilities.
**Breadth of Evaluation.** Our chosen axes of evaluation were not exhaustive and their interpretation was often subjective in nature. Although we conducted evaluations from both clinician and lay-perspectives, generating scenario-packs in three countries with assessors in both North America and India, the pool of clinicians and lay-people assessing the models could be expanded further to improve generalization of our insights. Our experiments could also undergo more extensive replication to explore other aspects such as inter-observer and inter-participant variability, including future work with an intentionally further diversified pool of human raters (clinicians and lay users). Participatory design in the development of model evaluation tools with a representative pool of patients, as well as clinical and health equity domain experts, could also be valuable.
Fig. 12: Experimental results on (a, b, c) Video-based recognition, (d) Image-based recognition, (e, f, g) Medical image-based segmentation.
**Note.** We provided annotators with some of the examples, information and tables in this document as part of the annotation guidelines. They are not repeated here for brevity.
We confirm the efficacy of the SLM in general pretraining by continual training Tinyllama-1.1B on 80 billion tokens. The results depicted in Figure 5 indicate that although Tinyllama has already undergone extensive training on the majority of these tokens, the application of SLM yields an average enhancement of 6.8% across 15 benchmarks compared to direct continual pretraining. The improvements were especially pronounced in code and math tasks, exceeding 10%.
**Evaluation of Fine-tuned LLMs.** Fine-tuned LLMs in this part refer to the model checkpoints obtained after instruction tuning or alignment tuning based on pre-trained model weights45.
* We conduct a study investigating the intrinsic tokenizer performance. * We conduct a study investigating the extrinsic tokenizer performance, i.e., the impact of the tokenizer on the model's downstream performance. * We investigate whether a correlation between the intrinsic and the extrinsic tokenizer performance exists.
We are grateful to the many researchers whose technical help, feedback, and discussions shaped this project: Jacob Austin, Samy Bengio, Olivier Bousquet, James Bradbury, Fernando Diaz, Mark Diaz, Noah Fiedel, Jonathan Frankle, David Grangier, Stefanie Karp, David Mimno, Gaurav Mishra, Michael Mozer, Sharan Narang, Alex Passos, Adam Roberts, Hanie Sedghi, Jascha Sohldickstein, David So, Florian Tramer, and Yun William Yu. We are also grateful to the Google Brain women who have given us continuous support.
To tackle the issue of inaccurate Internet retrieval results, Lazaridou et al. (2022) proposed using the TF-IDF algorithm to score the retrieval results. They used each question q verbatim as a query and issued a call to Google Search via the Google Search API. For each question, they retrieved the top 20 URLs and parsed their HTML content to extract clean text, generating a set of documents D for each question q. To prevent irrelevant information from hindering the resolution of a user's query, Hu et al. (2023) designed a gating circuit. This circuit utilised a dual-encoder dot product to calculate similarity and a gating circuit based on term weights. Additionally, Boytsov et al. (2016) presented an approach that replaced term-based retrieval with k-Nearest Neighbors(kNN) search while combining a translation model and BM25 to improve retrieval performance. This approach enabled the model to take into account the semantic relationships between terms and traditional statistical weighting schemes, resulting in a more efficient retrieval system.
Proof.: Consider any reward function \(r(x,y)\), which induces an optimal model \(\pi_{r}(y|x)\) under the KL-constrained RL problem, with solution given by 4.
By following these steps, it is possible to train an LLM using RLHF methods based on InstructGPT to create a human assistant that exceeds open assistant in performance, accuracy, and utility.
**Iterative self-supervision.** With a larger training corpus, our self-supervised fine-tuning approach can be used for iterative self-improvement. Specifically, after fine-tuning a model on its previous optimal predictions and reasonings, we can generate more fine-tuning data by using the same model again, which can be repeated until training data is exhausted.
We use "Answer the following questions with one or a list of entities. Do not give a detailed explanation. Answer needs to be as short as possible." for SQA. We find that OpenAI models often give detailed explanation for SQA dataset unless specifically instructed not to.
Figure 4: Data distribution of the 53 natural languages, aside from English,we used for pre-training. The number within each bar indicates the percent of the overall multilingual distribution that an individual language comprises.
- Intuitively, politicians often weigh their options carefully, and a decision to resign is not typically rushed, especially without a clear strategic reason.
**Wikipedia.** Wikipedia [159] is an online encyclopedia containing a large volume of high-quality articles on diverse topics. Most of these articles are composed in an expository style of writing (with supporting references), covering a wide range of languages and fields. Typically, the English-only filtered versions of Wikipedia are widely used in most LLMs (_e.g._, GPT-3 [55], LaMDA [68], and LLaMA [57]). Wikipedia is available in multiple languages, so it can be used in multilingual settings.
Overall, these downstream results validate that training a language model to predict the subtext between the lines of general text data can substantially improve its reasoning capabilities, even on datasets it was not explicitly trained on. The fact that longer rationales consistently lead to better outcomes, and that Quiet-STaR outperforms the constrained pause token approach, supports the notion that Quiet-STaR is successfully teaching the model to leverage its own generated thoughts to reason more thoroughly about the input.
As discussed in previous work [36], prompting is the major approach to utilizing LLMs for solving various tasks. Since the quality of prompts will largely influence the performance of LLMs in specific tasks, there have been a series of studies proposed to generate suitable task prompts through manual creation or automatic optimization, which will be introduced in this section.
We use the same parameters as Lee et al. (2022): \(n=5\) (5-grams); \(b=20\) and \(r=450\). This means that for each document, we compute a total of 9000 minhashes, and that the probability that a document pair with similarity 0.75 or 0.8 will be marked as duplicates will be \(76\%\) and \(99.4\%\) (respectively), diminishing rapidly for smaller similarity values.
**Training.** Similar to MiniCPM, MiniCPM-2.4B-128K utilizes the WSD as its learning rate scheduler and reuses the last checkpoint of the stable training stage of MiniCPM-2.4B. Concerning training data, we categorize the dataset distribution detailed in Section 6.3 into "short data" and "long data". We classify books, wikis, and papers as "long data", and the other as the "short data". The training comprises 44% long data and 56% short data for continued training. For the extension of long contexts, we apply Adjusted Base Frequency (ABF) (Xiong et al., 2023) in the 4K to 32k range and employ NTK-Aware RoPE Scaling (bloc97, 2023) and curriculum learning from 32K to 128K. Both two stages involves future training. Furthermore, as indicated in Yi Tech Report (Young et al., 2024) and Zebra (Song et al., 2023), we use of synthetic long QA data that significantly enhances model performance in context-aware tasks.
In this subsection, we establish the effect of increasing parameter count by an order of magnitude on the final performance of continual pre-training. To accomplish this we compare two continually pre-trained models to several baselines at 405M and 10B parameter model sizes in the _two dataset weak shift_ (Pile \(\rightarrow\) SlimPajama) and _two dataset strong shift_ (Pile \(\rightarrow\) German) settings.
where \(Z(x)=\sum_{y}\pi_{\text{ref}}(y\mid x)\exp\left(\frac{1}{\beta}r(x,y)\right)\) is the partition function. See Appendix A.1 for a complete derivation. Even if we use the MLE estimate \(r_{\phi}\) of the ground-truth reward function \(r^{*}\), it is still expensive to estimate the partition function \(Z(x)\)[17; 15], which makes this representation hard to utilize in practice. However, we can rearrange Eq. 4 to express the reward function in terms of its corresponding optimal policy \(\pi_{r}\), the reference policy \(\pi_{\text{ref}}\), and the unknown partition function \(Z(\cdot)\).
C-Evalis a comprehensive Chinese evaluation suite for foundation models. It consists of 13948 multi-choice questions spanning 52 diverse disciplines and four difficulty levels. We choosing the answer option with the highest log-likelihood as the final prediction of the model.
Lastly, we include an example from answering CommonsenseQA. Notably, this thought occurs while reading the question and hence was not used to predict the final answer.
We present a novel application of evolutionary algorithms to automate the creation of powerful foundation models. While model merging has emerged as a promising approach for LLM development due to its cost-effectiveness, it currently relies on human intuition and domain knowledge, limiting its potential. Here, we propose an evolutionary approach that overcomes this limitation by automatically discovering effective combinations of diverse open-source models, harnessing their collective intelligence without requiring extensive additional training data or compute. Our approach operates in both parameter space and data flow space, allowing for optimization beyond just the weights of the individual models. This approach even facilitates cross-domain merging, generating models like a Japanese LLM with Math reasoning capabilities. Surprisingly, our Japanese Math LLM achieved state-of-the-art performance on a variety of established Japanese LLM benchmarks, even surpassing models with significantly more parameters, despite not being explicitly trained for such tasks.
Xinyao Niu\({}^{4}\), Xinrun Du\({}^{1}\), Songyang Gao\({}^{5}\), Haoran Zhang\({}^{6}\),
The models are trained on A40 GPUs (48GB VRAM) for one epoch, taking up to 48 GPUs depending on the model size. The parameter-efficient training with LoRA is performed with PEFT library5. We also utilize DeepSpeed (Rasley et al., 2020) to optimize memory efficiency during the training process. We employ the AdamW optimizer (Loshchilov and Hutter, 2019) with a peak learning rate of 2e-4 and 5% warm-up cosine scheduler. Additionally, we apply gradient clipping with a value of 1.0 to mitigate potential gradient explosion.
With the emergence of the Mamba [12], the SSM model is increasingly attracting attention and favor from current researchers. The following works are currently explored in language modeling task [41][45][46][47][48], deep noise suppression task [42], and clinical note understanding task [49]. To be specific, for the language modeling task, [41] mainly studies the application of gated state spaces in the direction of long-range language modeling and introduces a novel method called GSS (Gated State Space). It can be used on long sequence modeling and effectively reduce the number of participants. Their experiments demonstrate that it achieves 2-3 times faster than the DSS [38]. Grazzi et al. [45] exploit the Mamba on simple function estimation and natural language processing in context learning tasks and validate that the overall performance is indeed better than the S4 version and comparable to other Transformer networks. S4++ [46] finds two issues of S4 architecture, i.e., the non-stationary state (NSS) and dependency bias, and proposes the State Memory Reply (SMR) mechanism to integrate multi-state information into the current state. They also integrate complex dependency bias via an interactive cross-attention mechanism and extensive experimental results show that S4++ outperforms S4 on multiple sequence modeling tasks, demonstrating significant performance gains.
Autonomous agents typically perform tasks by assuming specific roles, such as coders, teachers and domain experts [18, 19]. The profiling module aims to indicate the profiles of the agent roles, which are usually written into the prompt to influence the LLM behaviors. Agent profiles typically encompass basic information such as age, gender, and career [20], as well as psychology information, reflecting the personalities of the agent, and social information, detailing the relationships between agents [21]. The choice of information to profile the agent is largely determined by the specific application scenarios. For instance, if the application aims to study human cognitive process, then the psychology information becomes pivotal. After identifying the types of profile information, the next important problem is to create specific profiles for the agents. Existing literature commonly employs the following three strategies.
The LayerNorm operation is the composition of a centering operation \(x\mapsto x-\bar{x}\) with RMSNorm (i.e. first centering is applied, then RMSNorm). Therefore the identical argument to Appendix C.4 goes through, besides the fact that we need \(\mathbf{U}\) to also commute with the centering operation. Since the centering operation fixes a \((h-1)\) dimensional subspace defined by \(\mathbf{1}^{T}x=0\) where \(\mathbf{1}\in\mathds{R}^{h}\) is the vector of ones, it is enough to impose an additional condition that \(\mathbf{U}\,\mathbf{1}\in\{-\mathbf{1},\mathbf{1}\}\).
Start off by writing down sub-questions. Then use your sub-questions to help steer the search queries you produce.
In this chapter, we will list other works that represent early attempts at chain-of-thought reasoning or are designed for specific domains.
Figure 5: Layout of the survey in SurveyMonkey. Each respondent completed 25 similarly-formatted judgments.
As an example, the recruitment step allows agents to be removed or added based on the progress towards the goal. This helps ensure that the right agents are participating at any given stage of problem solving. The researchers found that horizontal teams are generally best suited for collaborative tasks like consulting, while vertical teams are better suited for tasks that require clearer isolation of responsibilities for tool calling.
**Quantized LLMs**. Compared with original models, quantized language models take a smaller memory footprint, and likely have a faster inference speed [427, 93, 413]. Recently, a number of quantized model copies of several publicly available language models have been released on HuggingFace, including BLOOM, GPT-J, and ChatGLM. In particular, GPTQ [417] has been widely used to quantize generative language models, leading to various quantized variants for LLMaM and OPT. Further, it has been also applied to quantize instruction-tuned models, such as Vicuna and WizardLM. Due to the large number of quantized LLMs, we do not directly incorporate the corresponding links of these models. The readers can easily find them by searching on HuggingFace.
\begin{table} \begin{tabular}{c c c c c} \hline \hline Location & \# of Scenario Packs & \# of Simulated Patients & \# of Patient Actors & \# of PCPs \\ \hline Canada & 60 & 67 & 10 & 10 \\ India & 75 & 82 & 10 & 10 \\ UK & 14 & 0 & 0 & 0 \\ **Total** & **149** & **149** & **20** & **20** \\ \hline \hline \end{tabular} \end{table} Table 1: **OSCE study summary.** Number of scenario packs, patient actors, simulated patients, and primary care physicians (PCPs) in each of the three locations (Canada, India, and the UK) in the remote OSCE study. 20 board-certified PCPs participated in the study as OSCE agents in comparison with AMIE, 10 each from India and Canada. 20 trained patient actors were involved, with 10 each from India and Canada. Indian patient actors played the roles in both India and UK scenario packs. Canadian patient actors participated in scenario packs for both Canada and the UK. This process resulted in 149 distinct simulated patients.
In addition to the general assessment of RALM, there has been some work focusing on the assessment of specific details and domains. RECALL (Liu et al., 2023) employs the EventKG and UJ datasets to incorporate inaccurate information into its existing data set. It then determines whether RALM is susceptible to being misled by such inaccurate information through two tasks: question answering and text generation. Xiong et al. (2024) concentrated on the medical domain and proposed MIRAGE, which integrates data from five datasets, including MMLU-Med, to evaluate the zero-shot learning, multi-choice evaluation, retrieval-augmented generation, and question-only retrieval ideation capabilities of medical RALMs. Ultimately, they also discovered the log-linear scaling property and the "lost-in-the-middle" effect in the medical domain.
Deactivating MethodTable 5 shows the result of the reasoning task, where we deactivate \(6\) sets of neurons. Details are listed in Appendix C.
These models use both encoder and decoder, and are sometimes called sequence-to-sequence models. At each stage, the attention layers of the encoder can access all the words in the initial sentence, whereas the attention layers of the decoder only accesses the words positioned before a given word in the input. These models are usually pre-trained using the objectives of encoder or decoder models, but usually involve something a bit more complex. For instance, some models are pretrained by replacing random spans of text (that can contain several words) with a single mask special word, and the objective is then to predict the text that this mask word replaces. Encoder-decoder models are best suited for tasks about generating new sentences conditioned on a given input, such as summarization, translation, or generative question answering.
In this section, we list some demonstrations of our selected multidisciplinary Chinese hard case instruction understanding and the following problem sets that are used in CHC-Bench. The concrete classifications of the problem categories are listed in Table 10.
Applications.Recently, LLM-based agents have shown great potential in autonomously solving complex tasks, making it feasible to rapidly develop capable applications for specific domains or tasks. In this section, we will discuss the applications in single-agent and multi-agent scenarios.
**Optimizer.** The Adam optimizer [328] and AdamW optimizer [329] are widely utilized for training LLMs (_e.g.,_ GPT-3), which are based on adaptive estimates of lower-order moments for first-order gradient-based optimization. Commonly, its hyper-parameters are set as follows: \(\beta_{1}=0.9\), \(\beta_{2}=0.95\) and \(\epsilon=10^{-8}\). Meanwhile, the Adafactor optimizer [330] has also been utilized in training LLMs (_e.g.,_ PaLM and T5), which is a variant of the Adam optimizer specially designed for conserving GPU memory during training. The hyper-parameters of the Adafactor optimizer are set as: \(\beta_{1}=0.9\) and \(\beta_{2}=1.0-k^{-0.8}\), where \(k\) denotes the number of training steps.
It is easy to find that the multiplier of the last and penultimate term is always \(\mathbf{C}\mathbf{\bar{A}}^{0}\mathbf{\bar{B}}\) and \(\mathbf{C}\mathbf{\bar{A}}^{1}\mathbf{\bar{B}}\). Therefore, we can treat these multipliers as the convolutional kernel \(\mathbf{\overline{K}}=\mathbf{C}\mathbf{\bar{B}}\cdot(\mathbf{\bar{A}}^{0}, \mathbf{\bar{A}}^{1},\mathbf{\bar{A}}^{2},...,\mathbf{\bar{A}}^{L})\), here, \(L\) is the length of the given input sequence. We can rewrite the Equ.
In the voting phase, we compute the BLEU score using _sacreBLEU_(Post, 2018) to evaluate the similarity between each of the generated samples. The sample with the highest cumulative BLEU score is selected as the final answer. This method ensures that the final output is the most representative and accurate piece of code as determined by consensus through similarity scoring among the samples.
We generate questions from six subject domains. To generate a large set of question-and-answer pairs, we extract a corpus of content webpages and then query GPT-4 to generate a question based on the text, along with the ground truth answer and the excerpt used to generate the question. For each dataset below, we provide the full prompts used to generate questions in the Appendix.
already deduplicated of exact matches during their creation, ExactSubstrdeduplication does not actually remove any examples.
LLMs are large-scale, pre-trained, statistical language models based on neural networks. The recent success of LLMs is an accumulation of decades of research and development of language models, which can be categorized into four waves that have different starting points and velocity: statistical language models, neural language models, pre-trained language models and LLMs.
In contrast, if 7/8 of data is bioS(\(N^{\prime}\)) with a very small \(N^{\prime}\), simulating highly repetitive knowledge appearing in training tokens (e.g., "da Vinci painted the Mona Lisa" in millions of variations), this may not affect the model's capacity for "standard" knowledge (e.g., those with 100 exposures):
As discussed in a recent study [493], LLMs exhibit the abilities of both task recognition and task learning in ICL, but the two abilities seem to be possessed with different model scales. As shown in the experiments [493], the ability of task recognition is easier to obtain, and even a small LM with only 350M parameters can exhibit this ability, while task learning can only emerge for LLMs with at least 66B parameters. Another study [499] also supports this finding with specially designed experiments. They set up the tasks with flipped and semantically unrelated labels in the experiment, which require task learning when performing ICL. The results suggest that small LMs tend to disregard the labels and mainly depend on their prior knowledge to accomplish the task, while LLMs have the ability to surpass their prior knowledge and acquire new knowledge from demonstrations, resulting in better outcomes. Furthermore, to improve the task learning ability, Meta-In-Context Learning [500] proposes to include multiple related tasks instead of just a single one in the prompt. In addition, Symbol Tuning [501] fine-tunes LLMs on demonstrations with semantically unrelated labels (_e.g.,_ foo/bar instead of positive/negative for sentiment analysis), forcing LLMs to learn the task from demonstrations instead of relying on prior knowledge.
Figure 4: Infini-Transformers obtain better Rouge overall scores with more book text provided as input.
Moreover, apart from the aforementioned findings, it should be mentioned that we plan to publicly release the intermediate checkpoints of Amber-7B and OpenLLaMA-7B, which not only enhances the comprehension of scaling laws but also helps to develop more effective training strategies for LLMs. In conclusion, we hope our findings and open-sourced checkpoints can guide the developers to understand the optimization process of LLMs and facilitate the growth of foundation models.
Tokenization referes to the process of converting a sequence of text into smaller parts, known as tokens. While the simplest tokenization tool simply chops text into tokens based on white space, most tokenization tools rely on a word dictionary. However, out-of-vocabulary (OOV) is a problem in this case because the tokenizer only knows words in its dictionary. To increase the coverage of dictionaries, popular tokenizers used for LLMs are based on sub-words, which can be combined to form a large number of words, including the words unseen in training data or words in different languages. In what follows, we describe three popular tokenizers.
Once the LLMs are trained, we can use them to generate desired outputs for a variety of tasks. LLMs can be used directly through basic prompting.
ChatQA-13B demonstrates better tabular QA capability given the scores in Avg-table. ChatQA-13B also shows better scores in datasets where documents that do not need retrieval, while it is on par with GPT-3.5-turbo in text-only documents and documents that need retrieval.
By examining these models across various stages of their training lifecycle, we aim to shed light on the dynamic interplay between model architecture, size, training strategies, and their impact on learning efficiency and task-specific performance. This comprehensive analysis not only assesses the current state of the art in language modeling but also explores the broader implications of model training and design choices on the future trajectory of large language model development.
Mistral-7B [65] is a 7B-parameter language model engineered for superior performance and efficiency. Mistral-7B outperforms the best open-source 13B model (LLaMA-2-13B) across all evaluated benchmarks, and the best open-source 34B model (LLaMA-34B) in reasoning, mathematics, and code generation. This model leverages grouped-query attention for faster inference, coupled with sliding window attention to effectively handle sequences of arbitrary length with a reduced inference cost.
Figure 33: **Llama Generated Data.** Llama2 finetuning when the loss for training and evaluation is the cross-entropy for all tokens in the chunks, including the prompt.
In this work, we take a step towards gaining a comprehensive understanding of the language capability transfer in LLMs. As shown in figure 1, we empirically investigate several key aspects based on LLaMA:
* Language models typically need at least 50K training steps _regardless of_ batch size. Thus, for small \(N\), we _reduce the batch size_ to ensure the total number of training steps exceeds this threshold. For very large models, a larger batch size is preferred to enable GPU parallelism. * When \(lr\) remains constant, \(wd\) should be relatively reduced as the number of training steps increases. Mathematically, the model weights should be "halved" for every \(\Theta(\frac{1}{lr\times wd})\) training steps. Therefore, it's advisable to reduce the \(wd\) parameter when training for longer periods.
**Better Prompt Design.** Since CoT prompting relies on prompts to elicit the reasoning capabilities of LLMs, the design of prompts is critical to its performance. As a direct approach, it is shown that using diverse CoTs (_i.e.,_ multiple reasoning paths for each problem) can effectively enhance the performance [437]. Another intuitive idea is that prompts with more complex reasoning paths are more likely to elicit the reasoning ability of LLMs [433], which can result in higher accuracy in generating correct answers. However, all these approaches rely on annotated CoT datasets, which limits their use in practice. To overcome this limitation, magic instructions such as _"Let's think step by step"_ can be used to automatically construct CoTs by prompting LLMs [434].
Observed from previous experiments and analysis, we can conclude that the pre-training loss is a good indicator of LMs' performance on downstream tasks, independent of model sizes, training tokens, languages, and pre-training frameworks.
SFT Data ScalingRecently, some research efforts have focused on the data scale for supervised fine-tuning. For instance, LIMA Zhou et al.
The most similar work to ours is the concurrent work of Guo et al. (2024). They employ various augmentations at the finetuning, rather than pre-training stage, including shuffling and reversing chunks of the input sentences. Unlike our method, their method first segments sentences in the training into semantically meaningful chunks via an LLM. While a chunk can be an entity name, it is more generally applied to all words, e.g. "of developing the first emotional" as a chunk. The actual segmentation is done via prompting another LLM with a specific instruction. Therefore, the unit of reversal will depend on the LLM and its prompt, making it presumably a difficult language modeling problem, whilst also requiring extra compute to reverse the sequence. This is applied only to finetuning on short sentences, which means the reversal curse mitigation is limited to the facts included in the finetuning data, and it is unclear if it can be applied to large pre-training documents. In contrast, our method is applied in the pre-training stage so it can learn to reverse a wide-range of general knowledge facts.
MMLU ResultsFor each task and model, we compared four approaches: using just the base model, RAG, FT, and finally combining FT and RAG by using the fine-tuned model as the generator. Furthermore, we tested the MMLU tasks using both 0-shot and 5-shot scenarios. The full results are shown in Table 1.
Task: Tell me if the following email is a promotion email or not. Class label: Promotion Email: Check out our amazing new sale! We've got discounts on all of your favorite products. Class label: Not Promotion Email: We hope you are doing well. Let us know if you need any help.
While LLMs are evaluated on a standard set of benchmarks designed to gauge their general understanding and reasoning capabilities, the benchmarks for agent evaluation vary greatly.
In addition, to reduce hallucinated answers in unanswerable cases, we aim to empower our model to explicitly indicate it when the answer cannot be found within the given context. To obtain these unanswerable data samples, we requested annotators to provide all context locations to the user question. Hence, it enabled us to construct unanswerable scenarios by deleting the text from the corresponding locations in the context. After deleting the relevant text to the question, we use a sentence, "Sorry. I cannot find the answer based on the context", as the response for the unanswerable questions. Finally, we construct another 1.5k user-agent turns with unanswerable annotations, which provides a good trade-off of answerable and unanswerable cases (see SS6.5 for details).
**Conclusions.** In Figure 8(b)-8(e), when junk data mimics random knowledge, capacity ratios are _significantly impacted_ unless training time is substantially increased. In Figure 8(f), if the junk data is highly repetitive, there is no degradation.
\(\bullet\)_Human-level test benchmarks_ aim to evaluate the comprehensive ability of LLMs with questions designed for testing humans, such as AGIEval [708], MMCU [709], M3KE [710], C-Eval [711] and Xiezhi [712]. These benchmarks encompass a wide range of domains, difficulty levels, and languages to provide a comprehensive evaluation of LLMs' general capabilities. Compared to publicly available models, models offering API services (_e.g._, GPT-4, ChatGPT, Claude) demonstrate superior performance compared to publicly available models on these evaluation benchmarks. As the best-performing model in evaluations, GPT-4 surpasses average human performance in AGIEval [708]. However, it still lags behind the top human performance on these challenging benchmarks. Hence, there remains ample room for further enhancements in the overall abilities of LLMs, particularly for publicly accessible models.
Both in terms of perplexity improvement at the end of training (see Figure A19) and efficiency (see Figure A18) we do not observe a significant difference between the OPT embedding space and embedding spaces M1 and M2 at the small model scale (125 million parameters). We note that M1 and M2 are significantly worse than the all-mp-net-base-v2 and all-mini-LM-v6 at small scales **and** suffer from the same problem of pruning away long documents (compared to the OPT embedding space), so we expect these models to under-perform the OPT embedding space at the 6.7b scale.
To assess the adaptation of SaulLM-7B backbone to the legal domain, we present perplexity scores across four document types: contracts, legal decisions, legislation, and party submissions. Refer to Figure 8 for the results. Our model, SaulLM-7B, consistently outperforms Mistral-7B across all categories, exhibiting lower average perplexity scores with reduced variance. Interestingly, Llama2-7B demonstrates lower perplexity specifically in legislation documents, suggesting a potentially higher proportion of legislative text in the pertaining corpora compared to Mistral-7B.
We show the details of pairwise comparison on Evol-Instruct benchmark with LLaMA-based models, as a demonstration of how CRR faithfully reflects the capability of the target LLMs trained by different methods. In Table 5, we observe that number of ties dominates the results and the number of wins are scarce. We attribute it to the fact that the target model is essentially distilling knowledge from the strong model. As a result, most of the time, the instruction-tuned target model is only able to respond as good as the strong model, through the lens of the LLM-based evaluator.
23. Iris Winnow pledges to stay by her brother’s side and continue her search for Roman Kitt.
\begin{table} \begin{tabular}{l} \hline \hline **Input** Tell me a bio about Ronaldinho (Bio)
We assess _WizardLM_, Alpaca, Vicuna, and ChatGPT on _Evol-Instruct_ testset and Vicuna testset using both automatic and human evaluations.
\begin{table} \begin{tabular}{l l l} \hline \hline **Category** & **Description** & **Number of links** \\ \hline adult & adult websites: from eroticism to hard pornography & 4516478 \\ phishing & phishing websites, malwares, etc. & 42445 \\ dating & dating websites & 3829 \\ gambling & online casino & 1365 \\ filehosting & websites hosting files, videos, pictures, music & 909 \\ ddos & websites related to ddos attacks & 421 \\ agressif & hate, racism, etc & 390 \\ chat & online chat websites & 244 \\ mixed adult & websites with some adult content & 153 \\ ariel & French regulated gambling websites & 69 \\ \hline \hline \end{tabular} \end{table} Table 13: **We select categories likely to contain adult or malicious content, as well as spam or unstructured text.**We curated the lists based on manual inspection of the data, informed by top hits reported by ToxicBERT. For the strict subword matching, we included words that were unequivocally related to adult content (e.g., groupsex). We avoided partial unclear matches (e.g., ass), that may be part of neutral words (e.g., massachusetts). In the soft word list, we included words that do not constitute a sufficient reason to discard the document on their own, but which are suspicious when multiple words from the list result in a match. This helped with keeping medical or legal content unaffected (e.g., a single match of dick).
Figure 1: An example of RAG benefits ChatGPT resolves questions that cannot be answered beyond the scope of the training data and generates correct results.
State-of-the-art LLMs continue to struggle with factual errors (Mallen et al., 2023; Min et al., 2023) despite their increased model and data scale (Ouyang et al., 2022). Retrieval-Augmented Generation (RAG) methods (Figure 1 left; Lewis et al. 2020; Guu et al. 2020) augment the input of LLMs with relevant retrieved passages, reducing factual errors in knowledge-intensive tasks (Ram et al., 2023; Asai et al., 2023a). However, these methods may hinder the versatility of LLMs or introduce unnecessary or off-topic passages that lead to low-quality generations (Shi et al., 2023) since they retrieve passages indiscriminately regardless of whether the factual grounding is helpful. Moreover, the output is not guaranteed to be consistent with retrieved relevant passages (Gao et al., 2023) since the models are not explicitly trained to leverage and follow facts from provided passages. This work introduces **Self-Reflective Retrieval-augmented Generation (Self-RAG)** to improve an LLM's generation quality, including its factual accuracy without hurting its versatility, via on-demand retrieval and self-reflection. We train an arbitrary LM in an end-to-end manner to learn to reflect on its own generation process given a task input by generating both task output and intermittent special tokens (i.e., _reflection tokens_). Reflection tokens are categorized into _retrieval_ and _critique_ tokens to indicate the need for retrieval and its generation quality respectively (Figure 1 right). In particular, given an input prompt and preceding generations, Self-RAG first determines if augmenting the continued generation with retrieved passages would be helpful. If so, it outputs a **retrieval** token that calls a retriever model on demand (Step 1). Subsequently, Self-RAG concurrently processes multiple retrieved passages, evaluating their relevance and then **generating** corresponding task outputs (Step 2). It then generates critique tokens to **criticize** its own output and choose best one (Step 3) in terms of factuality and overall quality. This process differs from conventional RAG (Figure 1 left), whichconsistently retrieves a fixed number of documents for generation regardless of the retrieval necessity (e.g., the bottom figure example does not require factual knowledge) and never second visits the generation quality. Moreover, Self-Rag provides citations for each segment with its self-assessment of whether the output is supported by the passage, leading to easier fact verification.
We parameterize the inversion model using the method described in Section 4 and select T5-base (Raffel et al., 2020) as our encoder-decoder backbone, which has \(222M\) parameters. We set the maximum sequence length to \(64\) for all experiments. We train models for \(100\) epochs with Adam optimizer with a learning rate of \(2e-4\). We use a constant learning rate with linear warmup over the first \(25,000\) training steps. We train in bfloat16 precision.
where \(x_{0}=0\) is the initial checkpoint and \(x_{n}=n\) is the final checkpoint. Substituting these into the fitted equation, we can obtain the Loss values at the start and end after fitting: \(\mathcal{L}_{\text{start}}=b\) and \(\mathcal{L}_{\text{end}}=an+b\). The change in loss can then be expressed as: \(\Delta\mathcal{L}=\mathcal{L}_{\text{end}}-\mathcal{L}_{\text{start}}\). Meanwhile, we represent the average Loss of the last checkpoint as \(\mathcal{L}_{\text{mean}}\).
Knowledge utilization is an important ability of intelligent systems to accomplish knowledge-intensive tasks (_e.g.,_ commonsense question answering and fact completion) based on supporting factual evidence. Concretely, it requires LLMs to properly utilize the rich factual knowledge from the pre-training corpus or retrieve external data when necessary. In particular, question answering (QA) and knowledge completion have been two commonly used tasks for evaluating this ability. According to the test tasks (question answering or knowledge completion) and evaluation settings (_with_ or _without_ external resources), we categorize existing knowledge utilization tasks into three types, namely closed-book QA, open-book QA43, and knowledge completion.
Repetition removal.Due to crawling errors and low-quality sources, many documents contain repeated sequences: this may cause pathological behavior in the final model (Holtzman et al., 2019). We could catch this content at the later deduplication stage, but it is cheaper and easier to catch it document-wise early on. We implement the heuristics of Rae et al. (2021), and remove any document with excessive line, paragraph, or n-gram repetitions.
Using **WRAP** on the C4, we evaluate model performance on 13 different zero-shot tasks, and 21 different language modeling domains of the Pile, and find that pre-training LLMs with synthetic data allows us to train equivalent models with 5x lesser data, or 3x lesser compute. In fact, our synthetic data trained models, also outperform the recent TinyLLama models that were trained for 3 trillion tokens (10x data and compute) across several zero-shot Q/A tasks. We further observe a reduction in perplexity by \(\sim\) 50% on the Pile, and note that our 350M parameter model trained on combinations of real and synthetic rephrases on just 15% of the entire C4 corpus, outperforms pre-training a 1.3B parameter on the entire C4. Finally, we conduct an analysis on the potential of data leakage, properties of synthetic data styles, and how to combine synthetic data for improving **WRAP** based LLM pre-training.
The learning rate scheduler (LRS), which adjusts the learning rate used in different stages of training, is crucial for model performance.
However, we found that this approach can generate inputs biased toward one label, especially for classification tasks (e.g., for grammar error detection, it usually generates grammatical input). Therefore, we additionally propose an **Output-first Approach** for classification tasks, where we first generate the possible class labels, and then condition the input generation on each class label. The prompting template is shown in Table 8.5 We apply the output-first approach to the classification tasks identified in the former step, and the input-first approach to the remaining non-classification tasks.
**Vignette Generator.** The vignette generator aimed to create varied and realistic patient scenarios at scale, which could be subsequently used as context for generating simulated doctor-patient dialogues thereby allowing AMIE to undergo a training process emulating exposure to a greater number of conditions and patient backgrounds. The patient vignette (scenario) included essential background information such as patient demographics, symptoms, past medical history, past surgical history, past social history, and patient questions, as well as an associated diagnosis and management plan.
This investigation into large language models (LLMs) enhances our understanding of model training complexities and scaling laws. Analyzing model dynamics across various tasks and training stages, we have gleaned insights to improve training and optimization strategies.
In our theoretical result, we introduce a dataset \(\mathsf{bioD}(N,K,C,D,L,T)\) defined as follows:
However, a few results have nuanced the view of scaling as a panacea to improved loss. For instance, (McKenzie et al., 2023) present evidence for "inverse sclaing" where flaws in the training objective or the data lead to U-shaped scaling.
Reinforcemnt Learning Another form of FT relies on RL or RL-inspired optimization strategies to better align the model after its pre-training phase. A few prominent examples are reinforcement learning from human feedback (RLHF) (OpenAI, 2023; Touvron et al., 2023), direct preference optimization (DPO) (Rafailov et al., 2023), and proximal policy optimization (PPO) (Schulman et al., 2017; Tunstall et al., 2023).
Figure 8: Distribution of the ROUGE-L scores between seed instructions and their most similar instructions in SuperNI (left) and the 252 user-oriented instructions (right).
1. Intensity: HIIT is more intense than other forms of exercise, which can lead to better results in a shorter amount of time.
The results are shown in Figure 3. When setting the linear coefficients of be [1, -1], we see that Yi 34B and LLaMA-2 70B performs the best in-terms of answer exact match. If we increase the number of the linear coefficients to be [1, 1, 1, 1, 1], we observe the emergent behavior that only large models (LLaMA-2 70B and Mixtral) can achieve good scores on exact match, although the differences to target is more continuous. These observations give side evidence for Yi-34B's performance on in-context learning and indicates that further scaling may allow the model to infer more complicated functions by in-context learning.
**Contribution on Experiments**. We would like to sincerely thank the following people for their hard work involved in experiments shown in Table 16.
A second scenario involves questions where prior knowledge about a topic may aid a model in answering.
We proposed ETS-DACP and ETA-DACP to enhance vanilla DACP by refining the pre-training data through active selection of relevant samples. We can select the data in two ways:Hard Sampling:We rank the samples in the domain corpus by the measure of choice. We select top-k samples from the domain corpus based on the metric(s), where \(k\) is the number of samples needed to hit the pre-decided token budget for continual pre-training.
tion of a reasoning process, such as Chain-of-Thought, to explain the provided answers.RAFT approach is similar: we demonstrate that creating a full reasoning chain and in addition, clearly citing sources enhances the model's accuracy in answering questions. In Fig. 3, we illustrate this set-up. Generating the training data in this fashion, involves presenting the model with a question, context, and verified answers, and then requesting it to form a reasoning chain that appropriately references the original context.
Fig. 8: Different selective scan methods used in SSMs proposed for image and video processing. (a) VMamba [61], (b) Vision Mamba [60], (c) RSMamba [139], (d-f) Video Mamba [84].
* Rating-A: The response is valid and satisfying. * Rating-B: The response is acceptable but has minor errors or imperfections. * Rating-C: The response is relevant and responds to the instruction, but it has significant errors in the content. For example, GPT3 might generate a valid output first, but continue to generate other irrelevant things. * Rating-D: The response is irrelevant or completely invalid.
Outside of the context of language, learning policies from preferences has been studied in both bandit and reinforcement learning settings, and several approaches have been proposed. Contextual bandit learning using preferences or rankings of actions, rather than rewards, is known as a contextual dueling bandit (CDB; [48; 14]). In the absence of absolute rewards, theoretical analysis of CDBs substitutes the notion of an optimal policy with a _von Neumann winner_, a policy whose expected win rate against _any_ other policy is at least 50% [14]. However, in the CDB setting, preference labels are given online, while in learning from human preferences, we typically learn from a fixed batch of offline preference-annotated action pairs [47]. Similarly, _preference-based RL_ (PbRL) learns from binary preferences generated by an _unknown_'scoring' function rather than rewards [9; 35]. Various algorithms for PbRL exist, including methods that can reuse off-policy preference data, but generally involve first explicitly estimating the latent scoring function (i.e. the reward model) and subsequently optimizing it [16; 9; 12; 34; 19]. We instead present a single stage policy learning approach that directly optimizes a policy to satisfy preferences.