Datasets:
Translation-EnKo
/
arxiv-translation

Dataset card Data Studio Files Community
2
text
stringlengths
4
4.47k
To test this, we assume a fixed token budget and a fixed data size which matches the token budget. We evaluate training on all the data as well as for two epochs on subsets of the data selected either randomly or using D4. We trained 1.3B parameter OPT models on these configurations and report average perplexity in Table 1. Unsurprisingly, epoching over a randomly selected subset of the data instead of using all the available data once leads to a slight degradation in model perplexity. In contrast, repeating data selected by D4 leads to an improvement in perplexity and downstream accuracy over randomly sampling new tokens. In other words, it is beneficial to select data via D4 and epoch 2 times, instead of doing one-pass learning on all available data. As seen in Figure 3, this finding generally holds across training as well. We refer to Section A.6 for results across model scale and data selection ratio.
In summary, our analysis of scaling laws, particularly through the scaling law, offers insights into training data size and model performance. We note improved performance with larger datasets, yet with diminishing returns. This indicates the importance of dataset expansion, alongside architectural and computational optimization, for maximizing data utility. Variations in scaling law efficacy across models suggest scaling behaviors are nuanced, necessitating tailored model development approaches. These insights hint at a future where refined scaling laws and optimization techniques significantly boost LLM capabilities, marking significant progress in AI.
This imposes a system of linear constraints on the logits. By querying the model many times, and accumulating many such systems of equations, we can recover the logit values more efficiently. To do this, we accumulate all such linear constraints in the set \(\mathcal{C}\), and at the end of each round, compute the smaller and largest possible values for \(\mathsf{logit}_{i}-\mathsf{logit}_{0}\) by solving a linear program that maximizes/minimizes this value over the constraint set \(\mathcal{C}\). Thus, at each round, we can maintain an interval that encloses \(\mathsf{logit}_{i}-\mathsf{logit}_{0}\), and refine the interval at each round given additional information from the queries made in that round. After \(T\) rounds (where \(T\) is chosen based on the total query budget for the attack), we return the tightest known bounds on each logit.
The effectiveness of these strategies is demonstrated in Table 8, which details the Yi-9B base model's performance across a variety of benchmarks, including common sense, reasoning, knowledge, coding, and mathematics. It underscores the competitive advantages of Yi-9B base model in specific domains, illustrating the efficacy of our methodology in enhancing model performance by optimally adjusting the interplay between data characteristics and model size.
**Pass@256 _v.s._ PassRatio@256 To deepen the understanding behind the performance improvements, we tracked Pass@N metric and PassRatio@N metric under different data size. The results are shown in Figure 3. With very limited synthetic data (_e.g._ 7.5K samples), the Xwin-Math-70B model already has very high Pass@256, indicating the strong ability to generate correct answers through multiple attempts. Meanwhile, the Pass@256 metric only changed slightly with increasing the amount of used data. In contrast, PassRatio@256, which reflects the stability to generate correct answer, increases significantly with the amount of synthetic data, and its growth trend is similar to that of Pass@1. This result confirms our hypothesize that the performance improvements is mainly caused by the better stability in answer generation rather than stronger ability to answer the question.
De-duplication refers to the process of removing duplicate instances or repeated occurrences of the same data in a dataset. Duplicate data points can introduce biases in the model training process and reduce the diversity, as the model may learn from the same examples multiple times, potentially leading to overfitting on those particular instances. Some works [125] have shown that de-duplication improves models' ability to generalize to new, unseen data.
**NOTE:** The model hallucinates information based on brief mentions of author's social media accounts.
**Internal Feedback.** The LLM itself can be utilized as a feedback provider. One straightforward way is to directly evaluate the quality of the generated plans through prompting. For example, RAP [447] evaluate the likelihood that each candidate plan can lead to task success, while Tree of Thoughts [527] proposes to vote across plans by making comparisons between them. Further, LLMs can provide feedback based on the intermediate results from the plan executor. For example, Reflexion [450] utilizes LLMs to transform sparse result signals (_e.g.,_ success or failure) into concrete text-based feedback (_e.g.,_ "_You should recommend comedies that the user mentions in the query instead of horror movies_") and stores this feedback in long-term memory for future planning.
The systems under test (SUTs) are general-purpose AI chat systems, which we define as AI models that have been trained (e.g., fine-tuned or instruction-tuned) to engage in open-ended conversations on a variety of topics. In principle, AI systems can have additional features, such as guardrails, safety-emphasising system prompts or safety filters.8 These could influence the systems' operations or how they are interacted with [30], but are not included with the SUTs for v0.5 for simplicity. We also do not test the Graphical User Interface or any other user-facing aspects that each system might have.
After data preprocessing, it is essential to design suitable strategies to schedule these multi-source data for pre-training a capable LLM. Generally, two key aspects should be paid close attention for data scheduling: the proportion of each data source (_data mixture_), and the order in which each data source is scheduled for training (_data curriculum_). Next, we discuss the two aspects in detail. An illustration of data scheduling has been presented in Figure 8.
* For \(N=100K\), we use \(wd=0.01\), \(lr=0.001/0.002/0.005\), and batch size 24 with fp16; * For \(N=200K\), we use \(wd=0.01\), \(lr=0.001/0.002/0.005\), and batch size 48 with fp16; * For \(N=500K\), we use \(wd=0.01\), \(lr=0.001/0.002\), and batch size 96 with fp16; * For \(N=1M\), we use \(wd=0.01\), \(lr=0.0005/0.001/0.002\), and batch size 192 with fp16; * For \(N=2M\), we use \(wd=0.005\), \(lr=0.0005/0.001\), and batch size 192 with fp16; * For \(N=5M\), we use \(wd=0.005\), \(lr=0.0003/0.0005/0.001\), and batch size 384 with fp16.
We use similarity between embeddings of task data and domain corpus samples to perform data selection. This allows us to select a subset from the domain corpus that closely resembles the distribution of task data. To quantify document-level task similarity, we employ cosine similarity between the document embedding and task data embedding. Prior works like [13] calculate embeddings from language model (RoBERTa) for a given unlabeled sample twice, which is not practical for LLMs. It takes a forward pass to compute the embeddings using LLM over entire corpus, or 25% of compute of using to train the pre-train the LLM over the entire corpus. We compute embeddings using the _Spacy_ model [9]. This approach allows us to cost-effectively measure the alignment between task-specific information and the financial corpus, enabling more focused and targeted pre-training.
The following are several passages about the demographics, symptoms, and management plan for a given condition. Generate 2 different patient vignettes consistent with these passages. Follow the format of the given example (just list N/A if a particular field is unavailable).
In Figures 1 and 2, we can separate the datasets into two groups: First, on TriviaQA, HellaSwag, RACE, WinoGrande, NLPCC-KBQA, ClozeT, CLUEWSC, and C3, the performance improves smoothly with decreased pre-training loss from the very beginning. Second, on MMLU, C-Eval, GSM8K, and GSM8K-Chinese, the performance remains flat when the loss is higher than a certain threshold. Once the pre-training loss is lower than this threshold, the performance starts to improve.
Figure 1: Overview of CodecLM. We first **encode** seed instructions into metadata to capture the underlying distribution of instructions. This metadata is then **decoded** through Self-Rubrics and Contrastive Filtering to tailor high-quality synthetic instructions that are aligned with the target instruction distribution. Intermediate instructions and responses are omitted in the figure for clarity.
Retrieval Aware Fine-Tuning (RAFT), presents a novel recipe to prepare fine-tuning data to tailor the models for domain specific open-book settings, equivalent to in-domain RAG In RAFT, we prepare the training data such that each data point contains a question (\(Q\)), a set of documents (\(D_{k}\)), and a corresponding Chain-of-though style answer (\(A^{*}\)) generated from one of the document (\(D^{*}\)). We differentiate between two types of documents: 'oracle' documents (\(D*\)) i.e. the documents from which the answer to the question can be deduced, and 'distractor' documents (\(D_{i}\)) that do not contain answer-relevant information. As an implementation detail, the 'oracle' document doesn't need to be a single document, but can be more than one document, as is the case in HotpotQA (Yang et al., 2018). Then, for \(P\) fraction of the questions (\(q_{i}\)) in the dataset, we retain the oracle document (\(d_{i}^{*}\)) along with distractor documents (\(d_{k-1}\)). For \((1-P)\) fraction of the questions (\(q_{i}\)) in the dataset, we include no oracle document and only include distractor documents (\(d_{k}\)). We then fine-tune the language model using the standard supervised training (SFT) technique, training it to generate answers from the provided documents and questions. Fig. 2 illustrates the high-level design principal for RAFT.
Stealing this layer is useful for several reasons. First, it reveals the _width_ of the transformer model, which is often correlated with its total parameter count. Second, it slightly reduces the degree to which the model is a complete "black-box", which so might be useful for future attacks. Third, while our attack recovers only a (relatively small) part of the entire model, the fact that it is at all possible to steal _any_ parameters of a production model is surprising, and raises concerns that extensions of this attack might be able to recover more information. Finally, recovering the model's last layer (and thus hidden dimension) may reveal more global information about the model, such as relative size differences between different models.
FLARE uses 5-shot in-context performance over the entire training data, _i.e.,_ each test example while evaluating each model sees different train examples.
Inspired by InstructGPT[2] and PRMs[36], we train two reward models to predict the quality of the instructions and the correctness of each step in the answer respectively:1. Instruction Reward Model (IRM): This model aims to judge the quality of the evolved instructions on three aspects: i) Definition, ii) Precision, and iii) Integrity. To produce the ranking list training data of IRM, for each instruction, we firstly use ChatGPT and Wizard-E 4 to generate 2-4 evolved instructions respectively. Then we leverage Wizard-E to rank the quality of those 4-8 instructions. Footnote 4: Wizard-E named Wizard-E-Generator, which is an Alpha version fine-tuned Llama model specifically used to execute _Evol-Instruct_ without APIs. 2. Process-supervised Reward Model (PRM): As there is no powerful open-source math reasoning LLMs before this work, there is no simple way to support highly precise process supervision without professional human-labelers and close-source ChatGPT. Therefore, we depend on ChatGPT to provide process supervision, and ask it to assess the correctness of each step in the solutions generated by our model. 3. PPO training. We evolve the original math (GSM8k + MATH) instructions by 8 turns, increasing the data size from 15k to 96k. We use IRM and PRM to generate the instruction reward (\(r^{I}\)) and the answer reward (\(r^{A}\)). Then apply a product as the final reward \(r=r^{I}\cdot r^{A}\).
\(\bullet\)_PagedAttention_. It has been observed when LLM are deployed on servers, GPU memory is largely occupied by cached attention key and value tensors (called _KV cache_). The major reason is that the input lengths are often varied, leading to fragmentation and over-reservation issues. Inspired by the classic paging technique in operating systems, PagedAttention has been proposed to improve the memory efficiency and throughput of deployed LLMs [285]. In detail, PagedAttention partitions each sequence into subsequences, and the corresponding KV caches of these subsequences are allocated into non-contiguous physical blocks. The paging technique increases the GPU utilization and enables efficient memory sharing in parallel sampling.
I want you act as a Prompt Rewriter. Your objective is to rewrite a given prompt into a more complex version to make those famous AI systems (e.g., ChatGPT and GPT4) a bit harder to handle. But the rewritten prompt must be reasonable and must be understood and responded by humans. Your rewriting cannot omit the non-text parts such as the table and code in #Given Prompt#:. Also, please do not omit the input in #Given Prompt#. You SHOUTLD complicate the given prompt using the following method: Please replace general concepts with more specific concepts. or You should try your best not to make the #Rewritten Prompt# become verbose, #Rewritten Prompt# can only add 10 to 20 words into #Given Prompt#.
Footnote 2: Most duplicates we saw were automatically generated pages, such as the outcomes of sports games. This shows the strength of manual curation for creating high-quality datasets.
providing an answer, and it replies that the context does not provide exact information about the age which is also correct.
Table 1 presents the results of this experiment, illustrating the performance of the Llama-2 model across the different configurations and datasets. Across both the NQ and SQuAD training datasets, the Adapt-LLM configuration consistently outperforms the Never Retrieve (NR-LLM) and Always Retrieve (AR-LLM) configurations on the PopQA test set. As can be observed, NR-LLM exhibits the lowest performance among the models, with an accuracy difference of approximately 14 absolute points compared to the other configurations. This disparity suggests that the parametric memory of Llama-2 alone is not sufficient for effectively answering PopQA questions.
For LLaMA-based models, we follow the practices in instruction tuning in prior works Zhou et al. (2023); Chen et al. (2023). We use AdamW optimizer with \(\beta_{1}=0.9,\beta_{2}=0.95\) to finetune the target model for 15 epochs, as suggested by Zhou et al. (2023) for smaller data size. We set the initial learning rate to \(1\times 10^{-5}\) and linearly decaying to \(1\times 10^{-6}\) by the end of training. We set per GPU batch size to 8, which is equivalent to a total batch size of 64, as we use 8 A100 GPUs for training. The maximum token length is set to 2048.
Jurning filter, lost, works on Plosing Time Arrays, a deep frequently wildly by Mrs. Price.
We have tried several {Instruction} for GPT-3.5-turbo and GPT-4, we find the ones below works the best (applied for both GPT-3.5-turbo and GPT-4).
Plan generation focuses on directly generating action sequences by prompting LLMs. Based on the format of the generated plans, existing work can be divided into two groups: text-based and code-based approaches.
The v1.0 benchmark will provide meaningful insights into the safety of AI systems. However, **the v0.5 benchmark should not be used to assess the safety of AI systems**. We have released it only to outline our approach to benchmarking, and to solicit feedback. For this reason, all the models we tested have been anonymized. We have sought to fully document the limitations, flaws, and challenges of the v0.5 benchmark in this paper, and we are actively looking for input from the community.
## 4 Nested scaling laws predict losses trained on various mixtures using only small-scale experiments
Let us denote by \(\mathcal{N}_{w,a}\) the set of \(n\in\mathcal{N}\) so that \(v^{\star}(n,a)=w\).
LLaMA/Mistral.Notably, as of the writing of this paper, LLaMA [32, 33] and Mistral [19] stand out as popular, publicly-available large language models. We highlight their key architecture differences from GPT2 -- which we define as having rotary embedding and no dropout.
Ablation studies.We conduct a set of ablations of our framework to identify which factors play key roles. We evaluate two model variants trained differently than our model: _No Retriever_ trains an LM using the standard instruction-following method given instruction-output pairs, without retrieved passages; _No Critic_ trains an LM trained with input-output pairs that are always augmented with the top one retrieved document without reflection tokens. This is similar to SAIL (Luo et al., 2023), and we use our instruction-output data instead of using the Alpaca dataset (Dubois et al., 2023), as in SAIL. We also conduct ablation on our inference-time algorithm, including _No retrieval_ disables retrieval during inference; _Hard constraints_ indicates the model performance that retrieves when \(\boxed{Retrieve}\)=Yes instead of using the adaptive threshold; _Retrieve top 1_ always retrieves and uses the top one document only, similar to standard RAG approaches; _Remove_ \(\boxed{Isstr}\) indicates the model performance that removes \(\boxed{Isstr}\) score only during critique-guided beam search in Eq. 4. In this ablation experiment, we use a training instance size of 50k for a more efficient exploration of training variations. Later in this section, we conduct an analysis of the effect of training data size. We conduct the ablation studies on three datasets, PopQA, PubHealth, and ASQA. On ASQA, we evaluate models on sampled 150 instances and exclude ablations involving adaptive or no retrieval processes.
Figure 2: **Tails of AI-generated data: Top.** Perplexity diagram of the Wikitext-103 test set, measured with Llama2–7B as the anchor model. We query the Wikitext-finetuned Llama2–7B to generate AI data, which is compared to the original set. Perplexity is calculated solely for the generated positions in both the AI and original datasets. AI data is generated for various settings of \((p,T)\). **Bottom.** Distribution of greatest common divisors (GCD) of pairs of random integers (original data (blue) scaling as \(p(GCD=k)\propto k^{-2}\)). A transformer is trained on this task on \(300M\) samples and used as a generator on a test set of randomly sampled integer pairs, giving the truncated GCD distribution.
2. Despite the advantages of a larger dataset, the improvements saturate. For example, RefinedWeb (320B) model outperforms the RefinedWeb (160B) model by only 0.2%. Similarly, the TinyLlama model (1T tokens) performs comparably to the **WRAP** model, which only had 85B tokens of raw web data.
Model merging in the PS and in the DFS are orthogonal approaches, however, it is straightforward to combine these disentangled methods and further boost the performance of a merged model. As we show in the rightmost illustration in Figure 1 and in Section 4.1, it is possible to first apply PS merging to a collection of models, and then put back this merged model in the collection and apply DFS merging from this enlarged collection.
Our test set comprises template-based QAs that feature single-entity answers. To generate QAs similar to those in the test set, we instructed the LLM to utilize a Chain of Thought (CoT) approach. However, we observed that not all the generated QAs adhered to our instructions.
Pretrained models are not always used as-is, but rather further fine-tuned to improve their performance, safety, and usability. Often models are first trained to follow instructions (Mishra et al., 2022; Wei et al., 2022; Sanh et al., 2022), and then further trained on human preferences (Ouyang et al., 2022) to improve the quality of their generations. We showcase the efficacy of using OLMo as a base model for further fine-tuning by training OLMo to be a general chat assistant following our Open Instruct (Tulu) data and training setup (Ivison et al., 2023). Our approach involves first performing instruction fine-tuning with a mixture of distilled and human-written instruction data and then further aligning the model with distilled preference data using Direct Preference Optimization (DPO) (Rafailov et al., 2023). We experimented with mixing the Tulu instruction data at the end of pretraining, as done in recent models such as DeepSeek-AI et al. (2024), but did not have conclusive findings.
1. Downward evolution: It enhances instructions by making the questions easier. For example i): revising high difficulty questions to lower difficulty, or ii) producing a new and easier question with another different topic. 2. Upward evolution: Derived from original Evol-Instruct method, it deepens and generates new and harder questions by i) adding more constraints, ii) concretizing, iii) increasing reasoning.
Figure 3: Recall@1 for four retrieval models across different popularity levels. The results indicate that retrieval models perform more effectively with less popular knowledge compared to more popular ones.
The generation stage is a crucial component of the RAG process, responsible for leveraging retrieved information to enhance the quality of the generated response. This stage encompasses several sub-steps aimed at producing content that is readable, engaging, and informative.
parameters, our \(\mathsf{bioS}(N)\) data utilizes only \(3275\) tokens (after applying GPT2's tokenizer), reducing the effective embedding layer size to \(3275\times(64h)\). This adjustment explains why, for \(\mathsf{bioS}\) data, GPT2small, typically known to have \(124\)M parameters, is counted as having only \(88\)M parameters in this paper.
Safety in PretrainingAligning with standard pretraining data safety practices [5; 58; 77], we build a set of filters based on heuristic rules, keyword matching, and learned classifiers to remove text containing personal identifiers and private data, and reduce sexual, violent, and extremist content.
Difference in revenue = $520 - $480 \\ Difference in revenue = $40 \\ \hline \end{tabular} \end{table} Table 4: A comparison case on different scale size modelsTherefore, the mechanic earned $40 more in revenue on the day with higher revenue, which is Thursday. The final answer is $40.
**Auto-evaluation of Conversation Ratings.** We leveraged the model-based self-CoT auto-evaluation strategy to rate conversations on four evaluation axes from the PACES rubric, and validated that these auto-evaluation ratings were accurate and well aligned with the specialist ratings (Figures A.17 and A.18). Furthermore, to demonstrate that the inner self-play loop improved simulated dialogue quality, we applied the auto-evaluation method to the simulated dialogues generated before and after the self-play procedure. Results in Figure A.19 revealed that the simulated dialogues after self-play were preferred more often than the baseline dialogues without self-critique.
The overlap is roughly constant across different vocabulary sizes, and the total overlap tends to be rather low, despite being the identical algorithm only implemented by two different libraries. Consequently, the tokenizers produce different tokenized sequences, possibly affecting model training and downstream performance. Investigating the underlying reasons, the low overlap might be attributed to different configuration and pre-processing options in these libraries. Due to the larger thesaurus in multilingual documents, the overlap for the multilingual tokenizer is lower than for the English tokenizers.
As discussed before, language model is not a new technical concept specially for LLMs, but has evolved with the advance of artificial intelligence over the decades. Early language models mainly aim to model and generate text data, while latest language models (_e.g.,_ GPT-4) focus on complex task solving. From _language modeling_ to _task solving_, it is an important leap in scientific thinking, which is the key to understand the development of language models in the research history. From the perspective of task solving, the four generations of language models have exhibited different levels of model capacities. In Figure 2, we describe the evolution process of language models in terms of the task solving capacity. At first, statistical language models mainly assisted in some specific tasks (_e.g.,_ retrieval or speech tasks), in which the predicted or estimated probabilities can enhance the performance of task-specific approaches. Subsequently, neural language models focused on learning task-agnostic representations (_e.g.,_ features), aiming to reduce the efforts for human feature engineering. Furthermore, pre-trained language models learned context-aware representations that can be optimized according to downstream tasks. For the latest generation of language model, LLMs are enhanced by exploring the scaling effect on model capacity, which can be considered as general-purpose task solvers. To summarize, in the evolution process, the task scope that can be solved by language models have been greatly extended, and the task performance attained by language models have been significantly enhanced.
Evaluation methodAll evaluations were done by appending each of the multiple-choice options to the question, followed by passing the concatenation through the model to get a log probability score per option. The highest score was interpreted as the model's choice and used for accuracy calculation.
To equip LLaMA with enhanced Chinese understanding and generation capabilities, we propose to continue pre-training the LLaMA model with Chinese corpora. However, directly applying continual pre-training with Chinese corpora encounters several challenges. Firstly, the original LLaMA vocabulary covers less than a thousand Chinese characters, which is insufficient to encode general Chinese texts. Although the LLaMA tokenizer circumvents this issue by tokenizing unknown UTF-8 characters to bytes, this strategy significantly extends sequence length and slows down the encoding and decoding efficiency of Chinese texts, as each Chinese character splits into 3-4 byte tokens. Secondly, byte tokens are not exclusively designed to represent Chinese characters. Since byte tokens also signify UTF-8 tokens in other languages, it becomes challenging for byte tokens and transformer encoders to effectively learn representations capturing the semantic meaning of Chinese characters.
**Training data and settings.** Our training data consists of diverse instruction-following input-output pairs. In particular, we sample instances from Open-Instruct processed data (Wang et al., 2023) and knowledge-intensive datasets (Petroni et al., 2021; Stelmakh et al., 2022; Mihaylov et al., 2018). In total, we use 150k instruction-output pairs. We use Llama 7B and 13B (Touvron et al., 2023) as our generator base LM, and we use Llama2 7B as our base critic LM. For the retriever model \(\mathcal{R}\), we use off-the-shelf Contriever-MS MARCO (Izacard et al., 2022) by default and retrieve up to ten documents for each input. More training details are in the Appendix Section B.1.
You must add [Shell cmd] format text as input data in [Rewritten Prompt]
Recently, a growing body of work has demonstrated the usefulness of custom tokens optimized to perform specific functions in the context of a neural network - for this reason, they have also been referred to as "function vectors." (Todd et al., 2023). One of the original instantiations of this was prompt-tuning (Lester et al., 2021) (and relatedly prefix-tuning (Li and Liang, 2021)), where the embeddings corresponding to the tokens of a prompt could be optimized to better accomplish a task. Others have applied meta-tokens to compress long prompts (Li et al., 2023; Jung and Kim, 2023) for efficiency. Most relevant to this work, Mu et al. (2024) optimized a token such that, when the tokens after it could not attend to the tokens before it (i.e., a context compression token), it would provide sufficient information to future tokens. Although we do not focus on compression, we share the problem of learning a token that affects attention and controls complex downstream behavior. In one related work, Goyal et al. (2023) show that learning a single "pause" token (essentially representing each token as two tokens) improves LM performance. However, unlike the thought tokens in our work, this pause token does not initialize a thought - instead, it can be seen as acting as the entirety of the thought. We find that reasoning in language is significantly more helpful.
Another limitation is enforcing the diversity in the generated data. This diversity comes from both the "style" and the "knowledge" contained in the generated data. Recent works (Li et al., 2023b;c) used a selection of topics, or scenarios to seed the model to generate novel texts. Still, a recent study by Padmakumar et al. (2023) showed that using language models for AI-assisted writing tends to reduce content diversity, particularly when using instruction-tuned models. While we used the paradigm of rephrasing specifically to mitigate the issues pertaining to the diversity of novel content generation, it remains for future work to assess the presence (or lack of) and impact of content diversity in paraphrase models.
Besides the above methods, there is extensive research on efficient tuning of Transformer language models. However, a more comprehensive discussion of efficient tuning is beyond the scope of this article, which can be found in the related papers on this topic [400, 408].
Original ResultsWe first evaluate the vanilla performance of Vicuna and Mistral on these datasets for later comparison as presented in Table 3.
\(\bullet\)_Carefully designed prompts can boost the zero-shot or few-shot performance of ChatGPT._ By comparing the results of using different prompts on the same task, we can see that using the carefully designed prompts can achieve better performance than the simpler ones. In the carefully designed prompts, we provide a more clearly expressed task description (_e.g._, WMT and WikiFact), or use a model-friendly format (_e.g._, GSM8k and OBQA). For example, for WikiFact task, the prompt with a more detailed task description leads to a performance increase from 29.25 to 31.21.
Adult user intent on malicious activities, behaving in a non-sophisticated wayThis persona is malicious, in the sense that they have intent or willingness to inflict harm. Specifically, they possess limited domain knowledge on how to cause harm, which results in them asking basic questions about harmful actions. They also have minimal understanding of AImodels and how their safety measures could be compromised. As a result, they do not employ advanced techniques such as jailbreaking or prompt injection.
With the advancement of Internet search and sorting technology, some researchers have focused their search efforts on Internet retrieval, which is a plug-and-play approach. This approach allows non-specialists to benefit from RALM and is better suited to the open domain and generalisation. Another advantage of this retrieval model is that it does not require real-time updating of the database, but relies on updates from commercial search engines. However, despite the advantages of simplicity and convenience, there is a significant amount of irrelevant and even harmful information on the Internet that can hinder the work of RALM. If an effective screening mechanism is not implemented, the effectiveness of RALM will be significantly reduced.
Meanwhile, the use of synthetic data has gained prominence in the paradigm of aligning pre-trained LLMs via instruction fine-tuning, RLHF (Ouyang et al., 2022), and instruction backtranslation (Li et al., 2023b). Recently, in the context of pre-training, synthetic datawas used to generate datasets such as Tiny Stories (Eldan & Li, 2023) and Textbook quality synthetic data (Gunasekar et al., 2023; Li et al., 2023c). These were used to train smaller language models (like the Phi model family) that were as performant as larger language models on certain tasks. However, their data generation process stays largely opaque, and prohibitively expensive, requiring prompting a GPT-3.5 model for generating billions of tokens. Additionally, such data generation can create a large "knowledge bias" by specifically generating data pertaining to tasks that we want to perform well on. While synthetic data has shown promise, it is unclear if this is because of the higher quality nature of synthetic data, or because of strategic topic selection (Maini, 2023).
PersonA: Yes, urinalysis can be used as a screening and diagnostic tool for bladder cancer. One way to test for bladder cancer is to check for blood in the urine, which is called hematuria. This can be done during a urinalysis by examining the urine cells under a special microscope.
Footnote 12: The expert annotators are instructed to use [https://c4-search.apps.allenai.org/](https://c4-search.apps.allenai.org/) for searching through the C4 corpus.
In-Domain Test Loss PredictionFor the in-domain test set, we use the code data following the same distribution as that used in the training data (the code data comprises \(10\%\) of the full training data).
**Result 9** (Figure 7).: _MoE is nearly fully efficient in storing knowledge_, capable of leveraging all its parameters despite the sparsity constraint._
We thank Jian Zhang, Zhenting Wang, and Wenyue Hua for their valuable discussions and suggestions during the project.
**Lemma 4.2**.: _In the logit-API threat model, under the assumptions of Lemma 4.1:_ (i) _The method above recovers \(\tilde{\mathbf{W}}=\mathbf{W}\cdot\mathbf{G}\) for some \(\mathbf{G}\in\mathbb{R}^{h\times h}\);_ (ii) _With the additional assumption that \(g_{\theta}(p)\) is a transformer with residual connections, it is impossible to extract \(\mathbf{W}\) exactly._
Figure 8: Gradient statistics over the training of a 0.2B model using WSD LRS. The exponential decay stage begins at 8000 steps.
Figure 3: **Specialist-rated top-k diagnostic accuracy.** AMIE and PCPs top-k DDx accuracy are compared across 149 scenarios with respect to the ground truth diagnosis (**a**) and all diagnoses in the accepted differential (**b**). Bootstrapping (n=10,000) confirms all top-k differences between AMIE and PCP DDx accuracy are significant with \(p<0.05\) after FDR correction.
**Instruct + Answers**: perplexity on instruction-tuning data from OPT-IML [21], where each sample contains both the instruction-tuning prompt and the answer (in Figure A4 this is referred to as "prompts_with_answers."
In an effort to understand why SentenceTransformer embedding spaces perform worse at larger model scales, we qualitatively analyze the clusterings with each SentenceTransformer embedding space. We find that using D4 with "all-mp-net-base-v2" and "all-mini-lm-v6" disproportionately prunes long documents. We hypothesize that this is because sentence transformer models are trained and finetuned on actual sentence pairs, which very rarely saturate the max context length of the model. This might result in all "long" documents (or at least any input that is max-context-length size) seem out-of-distribution to the model. We guess that this results in long documents being clustered together, and therefore disproportionately affected during pruning. This might be especially relevant in domains like Wikipedia articles, where headers and introductions look semantically similar, but the actual content (past the first max-context-length tokens) is very different.
Data AugmentationData augmentation is a well-established method for enhancing the performance of language models and has been surveyed extensively (Shorten
In the initial stage, individuals often use methods for retrieval that involve matching of relevant content, such as the TF-IDF Ramos et al. (2003) and BM25 Robertson et al. (1995) algorithms, which are considered classic and effective.
The Mturk interface used during construction of UpdatedLAMA and NewLAMA, NewLAMA-Easy, and NewLAMA-Easy are shown in Figure 5, 6, and 7, respectively.
_How can we keep our prompts safe?_ Our experiments show that when sampling is enabled, we can reconstruct model probability distributions given enough queries to the model. The only foolproof way to protect prompts while providing users access to generate text from a language model is to disable top-logits access (output only text) and set temperature to 0.
Previous research (Hoffmann et al., 2022) has established that the magnitude of the dataset significantly influences the performance of large language models. Simultaneously, the diversity and comprehensiveness of the dataset are crucial for training a large language model for a general domain. Guided by the aforementioned principles and our emphasis on utilizing Chinese corpora for model training, we have developed a dataset encompassing 1,254.68 billion tokens. This dataset integrates Chinese, English, and code data, consisting of 840.48 billion Chinese tokens, 314.88 billion English tokens, and 99.3 billion code tokens. The dataset aggregates content from diverse sources, such as web documents from Common Crawl, scholarly articles, encyclopedias, and books. The precise distribution is detailed in the Figure.1. Our dataset contains around 110 billion duplicate tokens, mostly in English. Despite being duplicates, they are high quality and were intentionally used twice in training.
As discussed in Section 3, our findings mirror a well-studied effect in humans, wherein recall is harder in the backward direction than in the forward direction (Clair-Thompson and Allen, 2013; Thomas et al., 2003; Bireta et al., 2010; Li and Lewandowsky, 1995; Guitard et al., 2019). For example, Li and Lewandowsky (1995) show that changing the visual-spatial characteristics of participants' study material affects backward recall, but not forward recall. It has been claimed that the two recall directions depend on different mechanisms in humans (Li and Lewandowsky, 1995). Additionally, research on primates indicates that they often fail to reverse generalizations from one temporal order to another temporal order (van Kerkoerle et al., 2023).
We use the LM-Evaluation Harness (Gao et al., 2021) to evaluate Nemotron-4 15B across all aforementioned tasks. Table 3 showcases that Nemotron-4 15B achieves the strongest average performance on this diverse set of tasks.
**RL Fine-Tuning Phase**: During the RL phase, we use the learned reward function to provide feedback to the language model.
Figure 19: **S-shape “Smoothed Grokking”.** Bigram data with Hutter++ model, mixing clean data with AI generated data with ratio 50 to 50. The grokking line is smoothed in the probabilistic setting. Line 1, 2, 3 are generated by using 10,000, 1,000, and 100 data to train the generating model. Compared to Figure 17, we do not limit the number of accessible real data now. \(\beta=3/2\).
English. On the contrary, when we only deactivate language-specific neurons not in the task-solving layer, non-English is influenced more seriously than English. Moreover, eliminating interference from the feed-forward layer achieves better results, which verifies the function of attention structure in the task-solving layer.
In recent times, substantial advancements have been witnessed in large language models (LLMs), exemplified by ChatGPT, showcasing remarkable proficiency across a range of complex tasks. However, many mainstream LLMs (e.g. LLaMA) are pretrained on English-dominant corpus, which limits their performance in other non-English languages. In this paper, we focus on how to effectively transfer the capabilities of language generation and following instructions to a non-English language. To answer this question, we conduct an extensive empirical investigation based on LLaMA, accumulating over 1440 GPU hours. We analyze the impact of key factors such as vocabulary extension, further pretraining, and instruction tuning on transfer. To accurately assess the model's level of knowledge, we employ four widely used standardized testing benchmarks: C-Eval, MMLU, AGI-Eval, and GAOKAO-Bench. Furthermore, a comprehensive evaluation of the model's response quality is conducted, considering aspects such as accuracy, fluency, informativeness, logical coherence, and harmlessness, based on LLM-Eval, a benchmarks consisting instruction tasks from 17 diverse categories. Our evaluation results demonstrate that comparable performance to state-of-the-art transfer models can be achieved with less than \(1\%\) of the pretraining data, both in terms of knowledge alignment and response quality. Furthermore, the experimental outcomes across the thirteen low-resource languages also exhibit similar trends. We anticipate that the conclusions revealed by the experiments will aid the community in developing non-English LLMs.
These techniques have been shown to be very useful, especially when used in conjunction with instruction tuning. However, similarly to instruction tuning, these methods focus on the overall quality of the response and its expected behavior and not necessarily on its breadth of knowledge.
Covering these six categories allows us to thoroughly examine the models' strengths and weaknesses across a diverse set of tasks, ensuring their robustness and adaptability. This comprehensive evaluation strategy enables us to identify areas where our models excel in understanding complex language constructs, applying commonsense knowledge, retrieving factual information, comprehending and analyzing written texts, solving mathematical problems, and performing in exam-like conditions.
_Remark 1.4_.: **Our framework offers a principled playground to compare models.** This contrasts with traditional comparisons based on loss/perplexity, which can produce debatable conclusions.6 Controlled data also reveal more significant differences between models.7 Footnote 6: A model might achieve better perplexity by performing _much better_ on simpler data but slightly poorer on complex data, or by excelling in reasoning tasks but not in knowledge storage. Our results offer a more nuanced view: GatedMLP doesn’t affect frequently encountered knowledge (with 1000 exposures) but does impact moderately rare knowledge (with 100 exposures). * Section 8: How quantization affects model capacity. We applied GPTQ[10] to quantize models from the base scaling laws to int8 or int4. Surprisingly, * Result 8: Quantizing to int8 does not compromise model capacity (even for models on the boundary of 2bit/param); however, quantizing to int4 reduces capacity to 0.7bit/param.
no human labeling; could more of this generated data lead to better instruction-following ability? We conduct an analysis of the size of generated data by subsampling different numbers of instructions from the generated dataset, finetuning GPT3 on the sampled subsets, and evaluating how the resulting models perform on the 252 user-oriented instruction set. We conduct the same human evaluation as in SS4.4. Figure 7 presents the performance of GPT3SELF-INST models finetuned with different sizes of generated data. Overall, we see consistent improvement as we grow the data size. However, this improvement almost plateaus after 16K. This is inline with the data scaling experiments in Wang et al. (2022, Fig. 5). Interestingly, when evaluating on SuperNI we found the model's performance gain plateaus earlier at around hundreds of instructions. This may be due to the fact that the new generated data is distinct from typical NLP tasks in SuperNI, indicating that future research may benefit from using a combination of different instruction data for better performance on various types of tasks.
**Reference [2]** Despite the official adoption of Sunday as a day of rest by Constantine, the seven-day week and the nundial cycle continued to be used side-by-side until at least the Calendar of 354 and probably later... The fact that the canon had to be issued at all is an indication that adoption of Constantine's decree of 321 was still not universal...
**Diversity** captures the diversity of distributions of token types in the domain corpus. Diversity has been shown to be an effective feature in related research on curriculum learning in language modeling [28, 21]. We use part-of-speech (POS) tagging to get token types. Since entropy has been shown to be one of the best measures of diversity [3], we use **entropy** of POS tags [28] as our diversity measure.
The result is shown in Table 10, we can see that on the most advanced smartphone iPhone 15 Pro, the inference throughput is as high as 18 tokens per second. In other devices, the inference throughput is also acceptable.
**MoD:** In [92], Tay et al. presented a generalized and unified perspective for self-supervision in NLP and show how different pre-training objectives can be cast as one another and how interpolating between different objectives can be effective. They proposed Mixture-of-Denoisers (MoD), a pre-training objective that combines diverse pre-training paradigms together. This framework is known as Unifying Language Learning (UL2). An overview of UL2 pretraining paradigm is shown in Fig 21.
In Figure 3 we directly inspect this model's final hidden activation vector across \(10,000\) different model queries and perform SVD of the resulting activation matrix. We see that despite GPT-2 actually having 768 potential hidden neurons, there are only \(757\) different activation directions. Thus, while this model is _technically_ a 768 dimensional model, in practice it behaves as if it was a 757 (i.e, the rank of the embedding matrix is 757) dimensional model, and our attack has recovered this effective size.
All annotators consented to the use and publication of their annotations. The dataset excludes copyrighted texts, containing only annotations done on model-generated summary claims. Additionally, we ensured annotators received fair compensation for their contributions.
By processing C4 through MDR, we are able to obtain subsets of data which might slightly outperform RefinedWeb; this combines both the stringent filtering of C4 (e.g., strict NSFW word blocklist, 3-sentence span deduplication) with our own filters and deduplication. While such a combination results in rejection rates that would be unacceptable for our target of 3-6 trillions tokens, this represents an interesting perspective for shorter runs, which may be able to extract extremely high-quality subsets from large web datasets.
\(\bullet\)_Natural Languages_. In this format, memory information such as the agent behaviors and observations are directly described using raw natural language. This format possesses several strengths. Firstly, the memory information can be expressed in a flexible and understandable manner. Moreover, it retains rich semantic information that can provide comprehensive signals to guide agent behaviors. In the previous work, Reflexion [12] stores experiential feedback in natural language within a sliding window. Voyager [38] employs natural language descriptions to represent skills within the Minecraft game, which are directly stored in memory.
**llama2 \(\rightarrow\) CP \(\rightarrow\) FT \(+\) chat vector** Continual-pretraining LLaMA2 on the target language corpus, fine-tune on the fine-tuning dataset and then adding chat vector.
Despite the performance improvement in complex reasoning tasks, CoT prompting still suffers from problems like incorrect reasoning and instability. In this part, we first introduce how to design better CoT prompts and enhanced CoT generation strategies, and then introduce the extension of the basic chain structure of CoT. Figure 15 illustrates the evolution of representative CoT prompting strategies.
\(\bullet\)_Base modification._ LLMs are usually trained with a pre-set maximum training length, _e.g._, 4096 in Llama 2 [99]. However, wavelengths in certain dimensions of RoPE may exceed the training length for longer text [276], so that language models have not undergone sufficient training (_i.e._, a complete rotation cycle) on these dimensions. Thus, when we adapt LLMs to longer texts, the rotation angles for certain dimensions would be never seen in the training phase [292]. Given a fixed rotation angle \(t\cdot\theta_{i}\), a smaller basis \(\theta_{i}\) allows for a greater distance \(t\), _i.e._, enabling the modeling of longer texts [235, 276, 288]. According to the formula \(\theta_{i}=b^{-2(i-1)/d}\) in Equation 4, decreasing the basis can be achieved by increasing the value of the base. In addition, decreasing the base can also help re-scale the wavelengths of all dimensions below the training length, while it often needs continual pre-training to adapt the LLMs to long context windows [292].
**Question:** How did Matthew Belk, a National Weather Service meteorologist, describe the September 2023 northeastern U.S. floods?
The answer is: \(\framebox{\frac{19}{3}}\)**Incorrect**Wrong calculation**
Figure 17: A response of the fine-tuned model on the question “Will Kevin McCarthy resign from the House by Nov 30?” (resolved NO) at the last retrieval index. Inspecting the original prompt (not shown here), we notice that our retrieval system presents articles, published on November 30, that clearly state that no official announcement had yet been made about Kevin McCarthy’s potential resignation. On the day of November, 30, the community prediction had dropped to 1%, but our system hedges by claiming 20%.
Language Model Training DynamicsInvestigating the training dynamics of language models is essential for understanding their behavior throughout the training process. This research includes studying internal representations (Saphra and Lopez, 2018), the acquisition of linguistic knowledge (Choshen et al., 2021; Liu et al., 2021), and the phenomenon of grokking (Power et al., 2022).
VM-UNet [65] establishes a baseline as the first pure SSM-based model for medical image segmentation. It competes effectively on ISIC17, ISIC18, and Synapse datasets, offering insights for future SSM-based segmentation systems. Gong et al. [66] propose the nnMamba which combines CNNs' detailed feature extraction with SSMs' broad dependency modeling, excelling in 3D medical image tasks. It proposes the Mamba-In-Convolution with Channel-Spatial Siamese learning (MICCSS) block to model the long-range relationship of the voxels. The superior performance is gained in 3D segmentation, classification, and landmark detection across 6 datasets. LMa-UNet [85] is a novel Large Window-based Mamba U-shape Network, leveraging large windows for improved spatial modeling compared to CNNs and Transformers, maintaining efficiency with linear complexity. It introduces a hierarchical and bidirectional Mamba block to enhance global and local spatial modeling. Tang et al. [82] use triple state space to fuse features in spatial and channel dimensions, and residual blocks to extract dense context features. Kazi et al. [102] take advantage of Mamba-UNet and the lighter version of the Hierarchical Upsampling Network (HUNet), the local feature extraction ability of the convolutional neural network is combined with the remote dependency modeling ability of the State Space Model. RS3Mamba [108] propose a novel two-branch network called RS3Mamba, which introduces a novel visual state space (VSS) model, Mamba, to the task of semantic segmentation of remotely sensed imagery. RS3Mamba constructs an auxiliary branch using the VSS blocks to provide additional global information for the main branch and introduces a co-completion module (CCM) to augment and fuse features from the dual encoder. Hao et al. [106] introduce a 3D CBCT segmentation method for teeth, termed T-Mamba, which enhances spatial position preservation and feature enhancement in the frequency domain by fusing shared position coding and frequency-based features. T-Mamba is the first work that introduces frequency features into the visual mamba architecture. Zhu et al. [113] propose a new semantic segmentation framework Samba based on the Mamba architecture and design specifically for high-resolution remote sensing images. Samba demonstrates the effectiveness and potential of the Mamba architecture in semantic segmentation of remote sensing images, surpassing current state-of-the-art CNN and ViT-based methods. Ma et al. [108] propose a novel dual branch network called RS3Mamba,which utilizes VSS blocks to construct auxiliary branches, providing additional global information for convolutional based main branches. In addition, considering the feature differences between the two branches, a Collaborative Completion Module (CCM) is introduced to enhance and fuse features from the dual encoder. Archit et al. [119] propose a new medical image segmentation network architecture, ViM-UNet. It is based on the latest Vision Mamba architecture and compared with traditional UNet and Transformer-based UNETR.