Datasets:
Translation-EnKo
/
arxiv-translation

Dataset card Data Studio Files Community
2
text
stringlengths
4
4.47k
\(\bullet\)_Increasing the topic diversity._ In addition to the complexity, improving the topic diversity of the instruction dataset can help elicit different abilities of LLMs on diverse tasks in real world [347]. However, it is difficult to directly control the self-instruct process for generating diverse instructions. Following YuLan-Chat [352], we employ ChatGPT to rewrite the instructions from Self-Instruct-52K dataset for adapting them into 293 topics via specific prompts. Finally, we obtain 70K instructions as the diversity-increased dataset.
For a period of time following the proposal of the retrieval technique, sparse retrieval proves to be a straightforward and effective tool in solving problems, particularly those based on knowledge. One of the main advantages of sparse retrieval is its simplicity, which can be easily integrated into existing indexing systems due to the fewer dimensions involved. (Hambarde and Proenca, 2023)This is consistent with human cognitive processes. Additionally, sparse retrieval is easier to generalise and more efficient.
Results.Figure 6 shows the performance of GPT3 model and its instruction-tuned counterparts on this newly written instruction set (w. inter-rater agreement \(\kappa=0.57\) on the 4-class categorical scale, see Appendix B for details). As anticipated, the vanilla GPT3 LM is largely unable to respond to instructions, and all instruction-tuned models demonstrate comparatively higher performance. Nonetheless, GPT3SELF-INST (i.e., GPT3 model finetuned with Self-INstruct) outperforms those counterparts trained on T0 or SuperNI data by a large margin, demonstrating the value of the generated data despite the noise. Compared with InstructGPT001, GPT3SELF-INST is quite close in performance--if we count acceptable response with minor imperfections (Rating-B) as valid, GPT3SELF-INST is only 5% behind InstructGPT001. Lastly, our evaluation confirms the impressive instruction-following ability of InstructGPT002 and InstructGPT003. Although there are many factors behind this success, we conjecture that future work can largely benefit from improving the quality of our generated data by using human annotators or training a reward model to select better generations, similar to the algorithm used by Ouyang et al. (2022).
Along this direction, researchers have developed numerous promising models (see Figure 1 for an overview of this field), where the key idea is to equip LLMs with crucial human capabilities like memory and planning to make them behave like humans and complete various tasks effectively. Previously, these models were proposed independently, with limited efforts made to summarize and compare them holistically.
Language models give the probability of the next token conditioned on the tokens that came before it, i.e. \(\mathbf{v}=p(x_{T+1}\mid x_{1},...,x_{T};\theta)\), where \(\mathbf{v}\in\Delta^{|\mathcal{V}|-1}\) gives the probability of each of \(|\mathcal{V}|\) possible next tokens. Generally these models have relatively large vocabulary sizes; the vocabulary \(\mathcal{V}\) may contain tens or hundreds of thousands of elements.
* In the first stage, we jointly optimize (1) and (2). Ranking by relevance and setting \(k=15\) achieve the lowest average Brier score. * In the second stage, we optimize (3) the reasoning prompt. We identify the top 3 prompts out of 15 candidates to elicit 3 predictions from our base model in our system; see Figure 15 for the best one.
Footnote 10: This is partly because exact-match is an easier metric for names than for descriptions.
As shown in Figure 5, memory manager manages short-term memory within an agent's lifecycle, ensuring that data is stored and accessible only while the agent is active, either waiting for execution or during runtime. The current AIOS supports storing each agent's memory independently, each of which other agents have no direct access to, unless it is authorized by the access manager. More complicated memory mechanisms such as shared memory pools among agents or hierarchical caches can be considered and integrated into AIOS in the future. Compared with the storage manager introduced in the following, the memory manager enables rapid data retrieval and processing, facilitating swift responses to user queries and interactions without overburdening the storage of AIOS.
Our method is parallel to the above fine-tuning methods. after obtaining supervisory signals through different approaches, one can consider using this regularization method in the final optimization step to further aid in training.
7. Now, we have another instance of \(f\). Since \(f\) is odd, we have \(f(-x)=-f(x)\). So, we can rewrite the expression as \(-f(f(x))\).
Multi-modality instruction following.Downstream applications of multimodal LLMs require reasoning and instruction following capabilities. Such data are usually long-form question response pairs and are expensive for humans to create. LLaVA (Liu et al., 2024b) uses existing image captions to prompt GPT-4 (in text-only mode) for writing diverse and long-form prompt-answer pairs. During multimodal LLM training, images and prompts are used as input while the captions and bounding box information can be hidden. Besides image captions, other sources of image attribute information such as object bounding box (Zhao et al., 2023), OCR (Zhang et al., 2023d) and deredered charts (Carbune et al., 2024; Masry et al., 2023) can all fit into such as image attributes + text LLM rewriting synthetic data pipeline.
In all cases, we perform zero-shot evaluation using the rank classification approach popularized by Brown et al. (2020). Under this approach, candidate text completions (e.g., different multiple-choice options) are ranked by likelihood (usually normalized by some normalization factor), and prediction accuracy is reported. While Catwalk implements several common likelihood normalization strategies, including normalizing by number of tokens (per-token normalization) (Brown et al., 2020; Liang et al., 2022), by number of characters (per-character normalization) (Gao et al., 2023), as well as incorporating an answer's unconditional likelihood (Brown et al., 2020), we selected the normalization strategies for each dataset separately. Specifically, we used unconditional normalization for arc and openbookqa, per-token normalization for hellaswag, piqa, and winogrande and no normalization for boolq, and sciq (i.e., tasks formulated as single token prediction tasks).
Results averaged on the small-=+ aggregate of 6 tasks are presented in Table 4. We observe relatively strong performance of all web datasets compared to The Pile, showcasing that curation is not a silver bullet for performant language models. We find C4 to be a strong pretraining dataset, in line with the findings of Scao et al. (2022)-however, The Pile comparatively underperforms more in our benchmarks. The relatively disappointing results on OSCAR-22.01 may be due to the main version of the dataset being distributed without deduplication. Regarding RefinedWeb, both filtering and deduplication significantly improve performance.
[37] claim that emergent abilities proposed in [47] are mainly a mirage caused by nonlinear and discontinuos metrics. [49] also support the idea.
Figure 1: Our pipeline for collecting faithfulness annotations in book-length summarization (Tables). First, (a) we generate summaries through hierarchical merging. Next, (b) we prompt GPT-4 to extract decontextualized claims. Finally, (c) we conduct a human evaluation of these claims, requiring annotators to validate each claim and provide their reasoning and evidence for the assigned label.
**1M passkey retrieval benchmark.** We replaced the vanilla MHA in a 1B LLM with Infini-attention and continued to pre-train on inputs with length of 4K. The model was trained for 30K steps with batch size of 64 before fine-tuning on the passkey retrieval task (Mothashami and Jaggi, 2024).
\(\bullet\)_Model-based evaluation_. Since human-based evaluation is both expensive and time-consuming, some work has proposed leveraging powerful closed-source LLMs such as ChatGPT and GPT-4 as a surrogate for human evaluators [727, 729]. For example, AlpacaEval [729] collects a set of instructions and utilizes a capable LLM (_e.g.,_ GPT-4) as the judge to perform pair-wise comparisons against the reference outputs. Furthermore, MT-bench [727] collects a set of multi-turn questions for evaluation and improves the reliability of LLM-based evaluators through methods like ICL and CoT. Compared with human evaluators, LLMs such as ChatGPT and GPT-4 can achieve high agreement with humans, in both small-scale handcrafted and large-scale crowdsourced evaluation tasks. Despite this, these closed-source LLMs are limited in access and have the potential risk of data leakage. To address this, recent work [727] has explored fine-tuning open-source LLMs (_e.g.,_ Vicuna [138]) as model evaluators using scoring data from human evaluators, which has narrowed the gap with powerful closed-source LLMs (_e.g.,_ GPT-4).
Footnote 11: In two cases, Claude-3-Opus refused to merge two summaries, as they were affected by the extra information available in the front and back matter and did not constitute a logical story. We excluded these cases from this analysis.
Figure 19: Examples on translation. We notice that MiniCPM-2.4B-DPO performs better in translating the technical part, but both models are unnatural in expressing “just starting to learn Chinese.”. However, both model successfully understand the instruction with traps, and generate French despite we do not include French corpus intentionally.
The target model in each step of sDPO is also initialized with \(M_{t-1}\), the aligned model from the last step. This ensures that the final model in sDPO has undergone training with the same amount of data as the final model in DPO. On the other hand, one concern of such design choice is that it may become increasingly difficult to stabilize the training of the target model as the steps progress, since it has already undergone training with a decreasing learning rate schedule in the preceding steps. Thus, another option is to use the initial SFT base model \(S\) as the target model for all steps of sDPO.
In summary, lower perplexity does lead to stronger generalization when responding to questions, but it comes at the expense of forgetting previously acquired knowledge.
**Theorem 3.2** (Grokking with Tail Cutting).: _Consider a sample of size \(T\) of which a proportion \(\pi\) comes from the true distribution \(p\) and the remainder comes from a version \(p^{\prime}\) of \(p\) with its tail chopped off at rank \(k\).
For base models, we consider both common language models, i.e., LLaMA-2 7B/13B/70B/Mistral-7B, and math-specific models, such as Llemma-7B, to assess the generality of the proposed approach.
To discover the data mixing laws, we encounter two challenges posed by their characteristics.
\(\bullet\)_Training_. Due to the huge model size, it is very challenging to successfully train a capable LLM. Distributed training algorithms are needed to learn the network parameters of LLMs, in which various parallel strategies are often jointly utilized. To support distributed training, several optimization frameworks have been released to facilitate the implementation and deployment of parallel algorithms, such as DeepSpeed [74] and Megatron-LM [75, 76, 77]. Also, optimization tricks are also important for training stability and model performance, _e.g.,_ restart to overcome training loss spike [56] and mixed precision training [78]. More recently, GPT-4 [46] proposes to develop special infrastructure and optimization methods that reliably predict the performance of large models with much smaller models.
Brier Score for random guess baselines on the four tasks are 0.25, 0.25, 0.1875, and 0.16. Only the largest model surpasses the random guess baseline. This also supports the existence of emergent abilities.
We here details our methodology for constructing preference data \(\mathcal{D}\). This dataset is developed using the FLORES-200 data (both development and test sets) and encompasses the same language pairs as discussed in Section 2. For each language pair, the dataset comprises 2009 parallel sentences.
2. Increased endurance: HIIT can help athletes improve their endurance by increasing their ability to perform at high intensities for longer periods of time.
Based on the stage-1 format template, the LLM inputs in stage-2 adds {Context for Latest Question} and {Instruction} from User, as depicted below:
We leveraged OpenAI API's function calling feature to ensure that the model would response with a strict format.
\(\bullet\)_Auxiliary optimization objectives._ Besides the primary cross-entropy loss, several studies propose auxiliary training loss to enhance the learning from the alignment data. First, since the responses of each instruction can be scored by the reward model, the ranking loss can be used to train the model to preserve the ranking order of these responses. For example, RRHF [390] samples responses from multiple sources, including model-generated responses, such as those derived from the model itself, ChatGPT, and GPT-4, as well as human-written responses, spanning both high-quality and low-quality instances. To align with the scores from reward models, it further optimizes the ranking loss by encouraging the model to have a higher conditional log probability for the response with a higher ranking. SLiCPF [391] proposes to assess the similarity between model outputs and human preference via the distance in the latent space, and introduces specific calibration and regularization loss to calibrate the candidate sequences based on human-preference data. Second, to enhance the relatedness between the response and the instruction, some work adopts contrastive learning to push up the probability of correct instruction-response pairs while pushing down incorrect instruction-response pairs. Specifically, for an output response, the proposed approach in [392] contrasts the target instruction to the other irrelevant instructions. By doing so, it can enable the model to learn the right correlation between instructions and responses.
The C4 dataset is a large, cleaned version of Common Crawl's web crawl corpus [18]. The context window was set to 1024, and each batch contained about 500k tokens.
The aforementioned findings enable us to transfer LLaMA's capabilities of language generation and following instructions to non-English languages at minimal cost. Based on evaluation results from four widely used standardized testing benchmarks (C-Eval, GAOKAOBench, MMLU, AGI-Eval) and an instruction evaluation benchmark LLM-Eval, we achieve comparable knowledge level and response quality to the state-of-the-art Open Chinese LLaMA, while using less than \(1\%\) of the training data. Furthermore, extension experiments on another 13 low-resource languages also exhibit similar trends. We aim for the experimental results and analyses in this paper to provide assistance and guidance to the community in constructing non-English LLMs.
The above attack makes a significant assumption: that the adversary can directly observe the complete logit vector for each input. In practice, this is not true: no production model we are aware of provides such an API. Instead, for example, they provide a way for users to get the top-\(K\) (by logit) token log probabilities. In this section we address this challenge.
We extend special gratitude to the Upwork annotators for their hard work, and to members from the UMass NLP lab for their feedback. This project was partially supported by awards IIS-2202506 and IIS-2312949 from the National Science Foundation (NSF) as well as an award from Adobe.
Ensure the new question only asks for one thing, be reasonable, be based on the Given Question, and can be answered with only a number (float or integer). For example, DO NOT ask, 'what is the amount of A, B and C?'.
\begin{table} \begin{tabular}{l} \hline **MODEL ANSWER:** Henry traveled 60-20= <<60-20=40>>40 miles between his first and second stops. He traveled 40 - 15 = <<40-15=25>>25 miles between his second stop and the end of the trip.
Reference ModelingWe begin by curating a high-quality dataset that reflects the desired data distribution. We train a reference model (RM) using standard cross-entropy loss on the curated data. The resulting RM is then used to assess the token loss within a larger pretraining corpus. We compute the reference loss (\(\mathcal{L}_{\text{ref}}\)) of a token \(x_{i}\) based on the probability that the RM assigns to this token.
For the \(100\)-exposure setting, we evaluate a broader range of model sizes per dataset. This approach is not only due to the tenfold reduction in training time compared to the \(1000\)-exposure setting but also to facilitate a detailed comparison of model architectures in the \(100\)-exposure setting, aiming for precision at higher model sizes.
Variation across PersonasFor many SUTs, we observe strong differences in the percentage of unsafe responses across the three personas (i.e., typical, malicious and vulnerable users). SUTs respond unsafely more to test items associated with _malicious_ or _vulnerable_ user personas than those associated with the _typical_ user persona. This trend holds across most hazard categories and SUTs.
Figure 2: Distribution of sample counts across buckets, defined by \(log_{10}\)(pageviews). The leftmost bin includes entities with fewer than \(10^{2}\) pageviews, while the right-most bin encompasses entities with over \(10^{5}\) pageviews.
In this section, we introduce our method which is implemented through a two-phase process: sampling and voting. The overview of our method is shown in Figure 2.
As introduced in Section 1, if each knowledge piece is seen 1000 times during training, we call this 1000 exposures. For \(\text{bioS}(N)\), 1000 exposures will unlikely include identical biography databecause there are 50 sentence templates for each attribute and a total of \(50^{6}\times 6!\) possible biographies per person. For \(\mathsf{bioS^{simple}}(N)\), 1000 exposures mean 1000 passes of the data. For \(\mathsf{bioR}(N)\), 1000/100 exposures mean only 25/2.5 passes of the training data.
DatasetFor testing, we used the MGSM dataset [41], a multilingual translation of a subset of the GSM8k dataset [7]. The Japanese test set of MGSM, consisting of 250 samples, was used for the final evaluation. We used a different dataset for evolutionary search to avoid overfitting the test set.
**Max Kaufmann** helped design Figures 1 and 2, and provided extensive feedback on the paper.
We believe that full access to open language models for the research community is critical to the scientific study of these models, their strengths and weaknesses, and their biases and risks. Accordingly, we introduce **OLMo**, a state-of-the-art, truly open language model and framework to build, study, and advance LMs, along with the training data, training and evaluation code, intermediate model checkpoints, and training logs.
Many language model benchmarks are designed to be solved in a single iteration, with no tool calls, such as MMLU or GSM8K [3; 10]. While these are important for measuring the abilities of base language models, they are not good proxies for agent capabilities because they do not account for agent systems' ability to reason over multiple steps or access outside information. StrategyQA improves upon this by assessing models' reasoning abilities over multiple steps, but the answers are limited to Yes/No responses [7]. As the industry continues to pivot towards agent focused use-cases additional measures will be needed to better assess the performance and generalizability of agents to tasks involving tools that extend beyond their training data.
URLs with many duplicates.Table 11 shows the URLs had the largest proportion of examples identified by NearDup as near-duplicates. For C4, these tend to be websites that sell many similar products and thus have a large amount of templated text. For RealNews, content aggregators seem especially common.
* 1 Introduction * 2 Main findings and Takeaways * 3 Related Work * 3.1 Continual learning * 3.2 Pre-training, Model Scale, and Continual Learning * 3.3 Domain Adaptive Continual Pre-training (DACPT) * 3.4 Continual Learning for LMs Applied to Specific Domains * 3.5 Learning Rate Schedules * 4 Background & Methodology * 4.1 Linear Warmup and Cosine Decay Schedule * 4.2 Compute-equivalent Replay * 5 Experimental Setup * 5.1 Datasets * 5.2 Continual Learning Settings * 5.3 Training Setup * 5.4 German and English LM Evaluation Benchmark * 6 Results * 6.1 Learning Rate Schedule * 6.1.1 The Effect of Linear Warmup for Weak and Strong Distribution Shifts. * 6.1.2 The effect of re-warming, re-decaying, and varying \(\eta_{max}\) for Weak and Strong Distribution Shifts. * 6.2 The Effect of Replay * 6.3 Continual Pre-training Final Performance for Weak and Strong Distribution Shifts.
We formalize this problem of prompt reconstruction as _language model inversion_, recovering the input prompt conditioned on the language model's next-token probabilities. Interestingly, work in computer vision has shown that probability predictions of image classifiers retain a surprising amount of detail (Dosovitskiy & Brox, 2016), so it is plausible that this also holds for language models. We propose an architecture that predicts prompts by"unrolling" the distribution vector into a sequence that can be processed effectively by a pretrained encoder-decoder language model. This method shows for the first time that language model predictions are mostly invertible: in many cases, we are able to recover very similar inputs to the original, sometimes getting the input text back exactly.
confidential data. Domain-adaptive continual pre-training (DACP) is a straightforward alternative to building from scratch; we continually pre-train a general-purpose LLM on a large scale corpus of domain-specific unlabeled data. Domain-adaptive continual pre-training has shown the ability to adapt the language models to better fit the in-domain distribution [8; 12; 32; 21]. They also enable large language models to acquire new knowledge as new data appears [11; 10], instead of training the model from scratch. We use DACP in our experiments to benchmark its benefits.
All of our released models have been trained to at least 2T tokens (a single epoch over our training data), and some have been trained beyond that by starting a second epoch over the data with a different shuffling order. The impact of repeating this small amount of data should be negligible according to prior work (Muennighoff et al., 2023).
Language model providers may be interested in defending prompts from inversion. One simple defense is to add noise to the language model output distribution; instead of providing a deterministic (argmax) output, from which an attacker could trivially reconstruct the output probabilities, language model providers, could instead _sample_ from the output distribution.
The Tree of Thought (ToT) [158] prompting technique is inspired by the concept of considering various alternative solutions or thought processes before converging on the most plausible one. ToT is based on the idea of branching out into multiple "thought trees" where each branch represents a different line of reasoning. This method allows the LLM to explore various possibilities and hypotheses, much like human cognitive processes where multiple scenarios are considered before determining the most likely one.
For summarization, we run our system end-to-end and pick the top 1 prompt (Figure 13) with respect to the Brier score.
First, we generate search queries that are used to invoke news APIs to retrieve historical articles. We initially implement a straightforward query expansion prompt (Figure 0(a)), instructing the model to create queriesbased on the question and its background. However, we find that this overlooks sub-considerations that often contribute to accurate forecasting. To achieve broader coverage, we prompt the model to decompose the forecasting question into sub-questions and use each to generate a search query (Min et al., 2019); see Figure 11(b) for the prompt. For instance, when forecasting election outcomes, the first approach searches directly for polling data, while the latter creates sub-questions that cover campaign finances, economic indicators, and geopolitical events. We combine both approaches for comprehensive coverage.
We propose a novel metric, **FUAR** (**F**rorgottes / (**U**pdated + **A**cquired) **R**atio), that can compare the efficiency of each CKL method using the trade-off between forgotten time-invariant knowledge and updated or newly acquired knowledge. FUAR represents relatively _how many_ time-invariant knowledge instances are forgotten in order to learn _one_ new or updated knowledge instance. We first define FUAR for the general case where there can be multiple corpora used for training an ever-changing LM.
Contributors to **license creation and risk assessment** include David Atkinson, Jesse Dodge, Jennifer Dumas, Crystal Nam, and Will Smith.
Rewritten Prompt(MUST contain a specific HTML page as input): I want to be able to scroll through the whole page, but without the scrollbar being shown.
The results are shown in Table 5. We sample multiple times from the models for each question and if any one of them contains the correct answer, then it is counted as success. The accuracy is relatively low in general due to the small model size in terms of number of parameters, limited pre-training and lack of any finetuning for both baselines and our method.
When evaluating the strengths and weaknesses of different ethical theories, it's important to consider their underlying assumptions, their implications for decision-making, and their ability to address the specific ethical dilemmas of a given context. It's also important to consider the practicality of implementing the theory and the degree to which it aligns with our own personal values and beliefs.
Planning can be flexibly combined with tools (Ruan et al., 2023) or agents (Crispino et al., 2023) to enrich reasoning ability. ToRA (Gou et al., 2023) designs mathematical specialized agents with external tools, and AutoUI (Zhang and Zhang, 2023) directly interacts with the multi-modal environment instead of converting visual inputs into text, which enhances the reasoning efficiency and reduces error propagation.
**Memory Formats**: In addition to the memory structure, another perspective to analyze the memory module is based on the formats of the memory storage medium, for example, natural language memory or embedding memory. Different memory formats possess distinct strengths and are suitable for various applications. In the following, we introduce several representative memory formats.
Evaluating the efficacy of LLMs is an evolving challenge in the field. In our study, we primarily employed tasks from the Huggingface Open LLM Leaderboard as benchmarks for evaluation. While this provided comparative results, future research could benefit from incorporating a wider array of tasks and benchmarks. These could include tasks that judge actual human or strong AI preference alignment. Such additional evaluation would not only enhance the validity of our findings but also contribute to the broader discourse on LLM assessment methodologies.
Figure 1: Training stages with parameter freezing. The fire and snowflake emojis indicate the trainable and frozen parameters respectively.
undergo two or more pre-training phases sequentially. At the start of each phase, we reset the optimizer states, since optimizer states may not always be available, e.g. when using open-weight models from HuggingFace. That is, our results apply to situations where a continually pre-trained LLM is randomly initialized and pre-trained on datasets \(\mathcal{D}_{0},\mathcal{D}_{1},\ldots,\mathcal{D}_{N-1}\) in sequence where \(N\geq 2\). For the realistic settings we consider _tokens_(\(\mathcal{D}_{i}\)) \(\geq\) 100B.
In this paper, we introduced a simple yet effective training method to help remedy the reversal curse in LLMs. Our reverse training works by first segmenting the input sequence into chunks and then reversing the ordering of chunks, but leaves the word-ordering in each chunk intact. A chunk can be a token, a word, an entity name, or a random number of tokens. The model is then trained on both the original sequences, and this reversed data. We evaluated on a symbolic reverse task and a reversing biography task that both demonstrated the necessity of preserving word-ordering within chunks. Next, we applied our reverse training to the realistic setting of LLM pre-training, which minimized the reversal curse on real-world knowledge. Evaluations on common benchmark tasks reveal that reverse training (particularly at the word level) during pre-training does not interfere with the forward prediction ability of LLMs, and actually improves metrics in the data-bound (rather than compute-bound) setting compared to standard training. When our method is applied to finetuning on fictitious facts, prediction accuracy rose from 0% to 70-100%. The reversal curse is a serious flaw in how LLMs acquire knowledge and reverse training opens a new promising direction in resolving it.
* _Workflow._ With the three components mentioned above, a typical workflow of an LLM-based agent is as follows. First, it receives information from the environment and writes it into short-term memory. Then, the agent processes the newly received information in the short-term memory. Such a process can be enhanced with information retrieved from long-term memory. Subsequently, the planning component utilizes the processed information from short-term memory to generate the next plan. Finally, the execution component carries out the plan generated from the planning component, which can be further assisted with external tools. By repeating the aforementioned process, the LLM-based agent can autonomously adjust its behavior in response to feedback from the environment and ultimately achieve its goal. Once LLM-based agents receive user requests or are assigned goals, they follow the above workflow to accomplish tasks through multi-turn interactions with the environment.
Some aspects of this work still warrant further research. For example, we focused on unsupervised training as our primary fine-tuning method, as opposed to instruction-tuning or RL-based methods. Researching combinations of various techniques, with diverse auxiliary knowledge bases, may yield improved results. This approach, combined with our hypothesis from Section 6, could further enhance our understanding of knowledge injection via FT.
**Reference [1]** Ronaldo Assis Moreira (born 21 March 1980), commonly known as Ronaldinho Gaichco or simply Ronaldinho, is a Brazilian retired professional footballer who played mostly as an attacking midfielder, but was also deployed as a winger. Considered one of the best players of his generation and regarded by many as one of the greatest of all time, Ronaldinho won two FIFA World Player of the Year awards and a Ballon d'Or. A global icon of the sport, he was renowned for his technical skills, creativity, dribbling ability and accuracy from free-kicks, as well as his use of tricks, feints, no-look passes and overhead...
Next, we will elaborate on our results and analysis in the following two subsections, illustrating the comparisons to the original LLaMA and other models.
Figure 1: Training loss for 1.4B models in the pre-training stage. On the \(x\)-axis we display the total number of tokens model has been trained on, including both in standard and reverse direction.
"chain of thought" prompting, where the model is guided to follow a logical reasoning process to arrive at an answer.
In this work, we introduce _Evol-Instruct_, a novel method using LLMs instead of humans to automatically mass-produce open-domain instructions of various difficulty levels, to improve the performance of LLMs. Figure 1 shows the running examples of _Evol-Instruct_. Starting from a simple initial instruction "1+1=7", our method randomly selects In-depth Evolving (blue direction line) or In-breadth Evolving (red direction line) to upgrade the simple instruction to a more complex one or create a new one (to increase diversity). The In-depth Evolving includes five types of operations: add constraints, deepening, concretizing, increase reasoning steps, and complicate input. The In-breadth Evolving is mutation, i.e., generating a completely new instruction based on the given instruction.
Figure 3: The impact of mask ratio when training on Llama2-7B with GSM8K. We compare MFT of two settings: Fixed mask ratio without warm-up, and a linear warm-up of the mask ratio starting from 0.
We deduce that \(c_{T}(i):=a_{T}(i)+b_{T}(i)\lesssim n_{T}(i)^{-c}\) for any \(i\). Importantly, the hidden constants don't depend on \(i\).
### Hoffmann et al.'s Approach 3 scaling policy is inconsistent with Chinchilla and our estimates
In LLMs, MoEs allow to train an extremely large model than is then only partially instantiated during inference when some of the experts are turned off wherever the gating/weighting function has a low weight assigned to them. As an example, the GLaM model has 1.2 trillion parameters, but during inference only 2 out of the 64 experts are used [84].
20. Iris’s journey is a search for her brother and Roman, and an exploration of human resilience in the face of conflict and divine manipulation.
Surprisingly, fine-tuning performs significant better than rewriting on E5-unsupervised. We conjecture that since E5-unsupervised does not use human-annotated query and context pairs in the pre-training stage, leading to a weak generalization to the high-quality rewritten query. In contrast, using high-quality dataset to fine-tune E5-unsupervised brings a giant boost with more than 15% improvements on both average top-1 and top-5 recall scores.
1. **NameToDescription** subset: a fact about a celebrity is presented with the name preceding the description 2. **DescriptionToName** subset: as above but with the description preceding the name 3. **"Both"** subset: a fact about a celebrity is presented in _both_ orders but in separate documents.
We make use of the two deduplication methods described in Lee et al. (2022): ExactSubstr and NearDedup (detailed in Appendix G.3.1 and Appendix G.3.2; see Appendix H for samples of duplicates).
Pythia and deduplication.The Pythia series of models is available in two flavours: one trained on the vanilla version of The Pile, and another trained on a version deduplicated with MinHash. Performance between these two flavours was noted to minimally differ (Biderman et al., 2023); in Figure 9, we find the deduplicated version may be slightly ahead of the non-deduplicated one under our aggregate. The higher end of this improvement is broadly in line with our findings in Table 5. Nevertheless, a difference in our findings and theirs remain.
**Finetuning for RAG** More recently, several papers have been exploring the idea of finetuning a pretrained LLM to be better at RAG tasks Lin et al. (2023); Wang et al. (2023); Xu et al. (2023); Liu et al. (2024). These works focus on constructing a combination of finetuning dataset for RAG and train a model to perform well on these tasks.
One can view the extra data \(\{\overleftarrow{x_{i}}\}\) as another language that the language model has to learn left-to-right - in this case a reversed natural language, which has a similar difficulty in terms of perplexity. As it is easy for the language model to identify which of these languages it is trying to generate from when predicting the next token, this does not tend to interfere with its language modeling abilities in the standard forward direction. Further, as it has been shown that LLMs can leverage knowledge across different sources (e.g., code to math (Shao et al., 2024), or different natural languages (Lample & Conneau, 2019)) we hypothesize that the knowledge it learns from the reverse direction can help in the forward direction as well.
RQ2: Does a combination of multiple synthetic datasets improve performance?We measure the impact of combining multiple synthetic styles with C4 for training. We consider two variants: combining in a 1:1 ratio meaning that there are two copies of C4 to match two synthetic styles (medium and QA), and 1:2 ratio which combines only one instance of the C4 dataset. For zero-shot QA tasks, our finding in Table 5-6 indicate lower performance than combining only QA and C4 data. Evaluations over the Pile are shown in Figure 4. We notice that both the 'Q/A' and 'Wikipedia' paraphrases help improve performance on certain domains. For example, 'Stackexchange', that has lots of question-answers benefits from the presence of synthetic data in Q/A style. Overall, we note that there is a small improvement on the average perplexity on the Pile by combining multiple styles.
GPT-1 [50] demonstrates for the first time that good performance over a wide range of natural language tasks can be obtained by Generative Pre-Training (GPT) of a decoder-only Transformer model on a diverse corpus of unlabeled text in a self-supervised learning fashion (i.e., next word/token predic
on 6144 TPU v4 chips using the Pathways system, which enables highly efficient training across multiple TPU Pods. PaLM demonstrates continued benefits of scaling by achieving state-of-the-art few-shot learning results on hundreds of language understanding and generation benchmarks. PaLM-540B outperforms not only state-of-the-art fine-tuned models on a suite of multi-step reasoning tasks, but also on par with humans on the recently released BIG-bench benchmark.
Besides, some researchers [312] propose to penalize the generation of previously generated tokens or \(n\)-grams to alleviate the issue of repetitive generation. In addition, diverse beam search [313] can be leveraged to produce a set of diverse outputs based on the same input.
* Hyperparameters such as batch size, learning rate, and learning rate scheduler influence the rate of convergence, yet do not impact the final converged loss provided that (1) their values fall within a reasonable range and (2) the model is trained with sufficient steps on adequate amounts of data. * Adjusting the batch size involves a trade-off between time and computation. The critical batch size that strikes an optimal time/computation balance can be determined based solely on the loss value. Training with this critical batch size requires twice as many training steps to achieve a specific loss value compared to using an infinite batch size (minimum possible required steps). * The context length, tokenization, data distribution and model configurations have big impacts on the constants in scaling law formulas, but do not affect the form of scaling law itself. * When given a fixed context length, tokenization, data distribution, model configurations and learning rate scheduler, we observe precise and predictable power-law scalings for performance in relation to training step, batch size, and model size, provided that the learning rate is optimally configured. * By training models with fewer than 60 million parameters, we can accurately estimate the constants in scaling-law formulas. This allows us to predict various attributes for models with up to 33 billion parameters before their training, including (1) the minimum possible loss; (2) the minimum required training steps and processed tokens to achieve a specific loss; (3) the critical batch size at any loss value; and (4) the complete test loss trajectory with arbitrary batch size. * These predicted attributes have many intriguing features, assisting us in identifying crucial factors before training large models, such as the optimal model size and training steps within a fixed computational budget, the necessary amount of data, the ideal mix ratio of multiple datasets, and more.
For the stage-1 SFT data used in stage-2 instruction tuning, we keep the same format template as in stage-1. In other words, the {Context for Latest Question} and {Instruction} are empty.
**Rank-order Agreement.** We evaluated our auto-evaluation method by quantifying its agreement with the specialist rankings of the OSCE dialogues. We limited our analysis to the 149 dialogue pairs in the study. Thus, each pair consisted of a AMIE conversation and a PCP conversation with the same patient actor, and rated by the same specialist. For a pair of two dialogues, the three possibilities were: the first one was rated better than the second one, they were equally rated, or the first one was rated worse than the second one. We defined the rank-order agreement as the proportion of dialogue pairs for which the specialist ranking was preserved by the auto-evaluation ratings. For example, we counted it as correct when our auto-evaluation rated AMIE's dialogue as better than the PCP's dialogue if specialists also rated AMIE's dialogue as better, regardless of the exact scores each method assigned.
Figure 5: Analyses for RAG preference rate against prior probability and deviation, using GPT-4 (blue), GPT-3.5 (orange), and Mistral-7B (green). Please see Figure 2 for full figure descriptions. Of note, some models did not generate any meaningful prior responses (due to refusal, improper responses, etc.) for certain datasets and thus could not be analyzed.
- There could be a significant underestimation of 'Oppenheimer's appeal to a specific demographic that turns out in large numbers.
* For presenting _negative_ results (Figure 5(a) and Figure 5(c)), we select the best learning rate from three options, identical to GPT2\({}_{0}\) in Parameter 4(c). * For presenting _positive_ results (Figure 5(b) and Figure 5(d)), we use a single set of learning rates, identical to Parameter 2 but with fp16 replaced by bf16 for a stronger comparison.
In the chain-of-thought reasoning scenario, where the demonstrations contain reasoning process, we need to maximize the likelihood of Answer \(\mathcal{A}\) and rationale \(\mathcal{R}\), as shown in Equ (3,4,5,6).
ObservationsBased on these two metrics, we examine the performance of the LLaMA-2 models on the GSM8K and the MATH benchmarks4 as shown in Figure 1. To adapt models for these two benchmarks in instruction-following settings, we use their SFT versions, which are trained with a limited amount of SFT data (i.e., 7.5K). As demonstrated in Bai et al. (2022); Ouyang et al. (2022), the SFT stage does not enhance capabilities (and may even lead to a reduction, as mentioned in the context of "alignment taxes"). Therefore, employing the SFT version provides a fair assessment of the models' mathematical capabilities.
find that the evidence often does not provide enough context that would allow someone who has not read the book to determine the faithfulness of the claim. As a result of decontestualization, claims always refer to people by name, but evidence often use pronouns instead. The annotator would need to quote a much larger chunk from the book in order for the evidence to include names as well. An even trickier case is that when dealing with high-level claims like "X is the protagonist of the story" or "The themes of the story are X, Y, and Z," one needs knowledge of the entire book, but citing the entire book as evidence is trivial. If annotators were to collect self-contained and sufficient evidence for every claim, the task would become significantly more challenging, sometimes even impossible. This difficulty with evidence gathering sheds light on why automatic evaluation does not work so well for this task.
Additionally, the experimental results indicate a correlation between the efficacy of the performance improvements and the difficulty of the problems addressed. To understand the reasons behind these performance improvements, we analyze the influence of problem difficulty on the effectiveness of our method. We classify difficulty into three dimensions: the inherent difficulty, the length of reasoning steps, and the prior probability of the correct answer. Through a series of experiments, we adjust these dimensions and observe their effects independently. We observe and summarize a few properties, based on which, we further develop optimization strategies that can intrigue the power of "More Agents".
To perform the experiment illustrated in Figure 2, we sample 100 paragraphs from Wikipedia obtained via the Wikitext dataset (Merity et al., 2016). We prompt GPT-4 with the first ten words of each paragraph prepended by the text _"Please update the sentence by replacing one word sentence with a close synonym. Respond with only the word to swap in the format word1 -? word2."_. We then extract the word swap from GPT-4's response and apply it to the input to produce the transformed input \(\hat{x}\). The language model used for prompting is the 7-billion parameter version of LLAMA-2 (non-chat version).