Datasets:

Translation-EnKo
/

arxiv-translation

Dataset card Viewer Files Files and versions Community

Split (1)

train · 6.9k rows

text stringlengths 4 4.47k
Previous studies have demonstrated that continual pre-training can significantly improve the model's ability to understand and generate specific content Gupta et al. (2023). However, continual pre-training could lead to catastrophic forgetting French (1999), and limited research has explored the abilities forgotten during pre-training on an existing fine-tuned LLM.
Conversational QA involves retrieval-augmented generation (RAG) in open-domain setting, or when the provided documents are longer than the context window of LLM. The dense retrievers are usually trained to retrieve the top-$k$ relevant chunks given a single question (e.g., Lin et al., 2023; Wang et al., 2022; Izacard et al., 2022). In conversational QA, the follow-up questions (e.g., with pronouns referring to entities mentioned in the previous conversation) may have insufficient information for retrieval, while feeding them along with the dialogue history can be redundant and lead to sub-optimal results.
RLHF System. The RLHF system mainly comprises three key components: a pre-trained LM to be aligned, a reward model learning from human feedback, and a RL algorithm training the LM. Specifically, the _pre-trained LM_ is typically a generative model that is initialized with existing pre-trained LM parameters. For example, OpenAI uses 175B GPT-3 for its first popular RLHF model, InstructGPT [66], and DeepMind uses the 280 billion parameter model Go-pher [64] for its GopherCite model [370]. Further, the _reward model (RM)_ provides (learned) guidance signals that reflect human preferences for the text generated by the LM, usually in the form of a scalar value. The reward model can take on two forms: a fine-tuned LM or a LM trained de novo using human preference data. Existing work typically employs reward models having a parameter scale different from that of the aligned LM [66, 370]. For example, OpenAI uses 6B GPT-3 and DeepMind uses 7B Gopher as the reward model, respectively. Finally, to optimize the pre-trained LM using the signal from the reward model, a specific _RL algorithm_ is designed for large-scale model tuning. Specifically, Proximal Policy Optimization (PPO) [128] is a widely used RL algorithm for alignment in existing work [66, 116, 370].
In our exploration of whether large language models (LLMs) should always be trained with the oracle context for Retrieval-Augmented Generation (RAG), we address a key question: what proportion (p%) of the training data should include oracle documents? Intuitively, one might assume that for effective training in reading and extracting information from context (e.g., RAG tasks), the oracle document should always be included during training (P = 100%). However, our findings challenge this assumption: incorporating a portion of the training data without the oracle document in the context (P = 80%) appears to enhance the model's performance on RAG tasks.
3. Battle Ropes: Battle ropes involve swinging a thick rope in a rhythmic motion, providing a full-body workout with high levels of intensity.
Natural language processing (NLP) field has witnessed a substantial paradigm shift with the advent of Large Language Models (LLMs). These models, distinguished by their considerable size and comprehensive training data, have demonstrated extraordinary abilities in comprehending and producing human-like text. In contrast to pre-trained language models dedicated to text understanding, such as BERT (Devlin et al., 2019), the GPT series (Radford et al., 2018) accentuates text generation, positioning them as more suitable platforms for creativity compared to their counterparts. Notably, the latest members of the GPT family, namely ChatGPT and GPT-4, have garnered significant attention, establishing themselves as leading examples in this rapidly evolving field.
Analysis. It is easy to see that the query cost of this attack is $\frac{1}{4m}$, where $m$ is the expansion factor. Further, since each query requires $1+m$ tokens, the token cost is $\frac{1+m}{4m}$. (Or, $1+m+\Delta$ if the API has an overhead of $\Delta$ tokens.) Note that if $m=1$, i.e., there is no expansion, this attack reduces to our first attack and the analysis similarly gives a query cost of $\nicefrac{{1}}{{4}}$ and a token cost of $\nicefrac{{1}}{{2}}$.
In Figure 8, we compare the results of the head bottleneck experiment of the Pythia-1B model in Section 5.1 to the $W$-error on the head of the same model as the bottleneck dimension $d$ evolves. It shows that the loss gap grows slowly with the $W$-error, implying that even when the allowed rank would lead to a poor approximation of $W$, the performance can still remain acceptable. We notice that the performance starts decreasing when the $W$-error outgrows 0.6.
Theoretical Models for Scaling Laws:From a theoretical angle, scaling laws have been shown analytically even before the emergence of large foundation models. For instance, Caponnetto and de Vito (2007) characterize the power-law generalization error of regularized least-squares kernel algorithms. The role of optimization can also be taken into account in this setting ((Nitanda and Suzuki, 2021)). In the nonparametric literature, for example (Schmidt-Hieber, 2017) and (Suzuki, 2019) derived the test error scaling of deep neural network in fitting certain target functions and (Bordelon et al., 2020) analyze spectral dependence.
In order to verify that our codebase could be used on both NVIDIA and AMD GPUs without any loss in performance, we trained models on two different clusters:* LUMI: Provided by the LUMI supercomputer,4 we used up to 256 nodes on this cluster, where each node consists of 4x AMD MI250X GPUs with 128GB of memory5 and 800Gbps of interconnect. Footnote 4: [https://www.lumi-supercomputer.eu](https://www.lumi-supercomputer.eu) * MosaicML: Provided by MosaicML.6 (Datalysts), we used 27 nodes on this cluster, where each node consists of 8x NVIDIA A100 GPUs with 40GB of memory and 800Gbps interconnect. Footnote 5: The MI250X is a dual-chip module, meaning in practice that each physical device consists of two logical devices, so each node has 8 logical GPU devices with 64GB of memory each.
Since an int8-based model has an absolute upper bound $R(F)\leq 8$ on capacity ratio, we have:
Fig. 1: [left gray sub-figure] Block diagram representation of the linear state-space equations (re-draw based on state-space representation); [right sub-figure] The formulation of widely used Mamba architecture (re-draw from [12]).
Key findings highlight the influence of training strategies, dataset quality, and architecture on LLM efficiency and effectiveness. While model size impacts learning outcomes, innovative methods enable smaller models to rival larger ones.
Here, {question} represents the current code instruction awaiting evolution, and {method} is the type of evolution. The five types we used are listed as follows:Add new constraints and requirements to the original problem, adding approximately 10 additional words.
The AI Safety taxonomy contains 13 hazard categories, 7 of which are in the v0.5 benchmark. The other 6 will be addressed in future versions of the benchmark. Each hazard category in the v0.5 benchmark has a single test set, which comprises multiple test cases.
performance still surpasses that of standard SFT (45.0, +3.4). Nonetheless, it exhibits lower performance compared to the computation of losses across all tokens (47.1, +5.5). Predicting the next tokens for masked positions poses a greater challenge than for unmasked ones, as they lack adjacent unmasked tokens from which to gather local information. This increased complexity ultimately improves the performance.
Where $w_{1:T}$ is sampled from a fixed SFT model $p_{\text{sft}}$ in the first stage, and the new data is added to the original data for the second SFT stage. While $p$ in Eq. (3) is updated online. Singh et al. (2023) also proposes to do this two-stage method for multiple iterations.
Multiple Generations. The mechanics of the proof of Theorem 2.4 apply in this setting. See Figure 12 in Appendix B illustrating that Equation (13) keeps holding for probabilistic data distributions.
The concept of emergent abilities in large language models (LLMs) draws researchers attention, as model sizes continue to scale. The phenomenon of "double descent," where a model's validation loss first worsens and then improves with increasing model complexity, was first proposed by Nakkiran et al. (2019). Furthermore, the "grokking" phenomenon, where models continue to improve even without a reduction in training loss, has been detailed by Power et al. (2022) and Murty et al. (2023), providing insight into the non-linear learning trajectories of LLMs.
In real-word forecasting competitions, forecasters do not have to make predictions on every question in the platform at every possible date. Instead, they typically make predictions on questions where they have expertise or interest in and at times that they choose. Therefore, it is natural to leverage our system's strengths and weaknesses and decide accordingly if we should forecast on a retrieval date $k$ for a question $q$.
We further demonstrate the data quality by comparing the speed of preference increase during data scaling. As is shown in Fig. 5, when compared with UltraChat [19] and its cleaned version UltraChat 200K, we see a clear tendency of performance improvements when scaling up the Yi data.
Recent research has extensively explored the emergent abilities of LLMs in downstream tasks. Wei et al. (2022) highlighted that emergent abilities are predominantly observed as models scale up, underscoring the significance of model size in achieving advanced capabilities. Schaeffer et al. (2023) contended that the performance of emergent abilities is highly dependent on the chosen evaluation metrics, suggesting a nuanced relationship between model performance and assessment methodologies. In our study, we concentrate on the emergent abilities of LLMs in downstream tasks, utilizing common evaluation frameworks as provided by Contributors (2023). Our aim is to conclude the dynamics of emergent abilities within the context of practical applications, contributing to a deeper understanding of how LLMs can performs through scaling and standard evaluation metrics.
Many of the existing benchmarks focus on the ability of Agent systems to reason over logic puzzles or video games [17]. While evaluating performance on these types of tasks can help get a sense of the reasoning capabilities of agent systems, it is unclear whether performance on these benchmarks translates to real-world performance. Specifically, real-world data can be noisy and cover a much wider breadth of topics that many common benchmarks lack.
CC-RecentNews-Small into two different splits by the published date of each article to simulate a setting where multiple CKL phases are needed, denoted as Small-P1 (05.2020 - 11.2020)) and Small-P2 (11.2020 - 04.2021).
Consequently, we learn that the loss associated with each token during training does not decrease smoothly like the overall loss; instead, there is a complex training dynamic among different tokens. If we can select the appropriate tokens for the model to focus on during training, we may be able to stabilize the trajectory of the model's training and enhance its efficiency.
set, we generated 256 responses and used the number of steps in the GSM8k test set's CoT annotations as the theoretical CoT steps. We draw the curve to show the relationship between the number of CoT reasoning steps and mean final answer accuracy and show the fitted curve based on Equation. 3. We test Xwin-Math-7B models with different synthetic data, and the results are shown in Figure 4. The solid line is fitted using all seven points and Table 4 shows the estimated single-step accuracy when using different amounts of data using all data points, and it can be seen that the single-step accuracy improve significantly with more data.
$\bullet$_Edit-based approaches._ For the above methods, gradient-based and RL-based tuning can be extremely computationally demanding for ever larger models, and may not be feasible for API-based model calls (_e.g.,_ ChatGPT). Therefore, another line of work aims to directly edit existing prompts based on the task performance. Specifically, GPS [469] borrows an idea from the genetic algorithm and proposes a genetic prompt search method that utilizes a language model (_i.e.,_ T5) to edit prompts by taking the cloze task form. In addition to model based edit methods, human-defined operations can be also employed for prompt editing [470], including delete, swap, paraphrase, and addition. Based on these operations, they iteratively edit the prompts and greedily search for the best prompt guided by the model performance on a small pool of examples.
paths. In fact, it can be used as a plug-in to further enhance CoT-based methods, as will be shown in our evaluations.
More generally, one can envision adding domain names (e.g., wikipedia.org) to every piece of the pretraining data. This would significantly enhance the model's knowledge capacities, because Result 12 demonstrates that language models can automatically detect which domains are rich in high-quality knowledge and prioritize learning from them. We emphasize that the model does not need any prior knowledge to identify which domains contain high-quality knowledge; this process is entirely autonomous.
Figure 11: Performance of Baichuan-7B, DeepSeek-7B, DeepSeek-67B, Amber-7B, OpenLLaMA-7B, and Yi-34B models across reading comprehension benchmarks.
* CoQA (Reddy et al., 2019) is a conversational QA dataset with each dialogue grounded on a short passage. The answers are generally short, and the passages cover a wide range of domains like children's stories, literature, mid/high school exams, news, Wikipedia. * DoQA (Campos et al., 2020) covers three domains: cooking, travel, and movies collected from active Stack Exchange4 forums. The dataset contains unanswerable cases where answers cannot be found within the given document. Footnote 4: [https://stackexchange.com/](https://stackexchange.com/) * ConvFinQA (Chen et al., 2022) is based on the Financial domain. Each document contains a single financial report table along with relevant text surrounding the table. This dataset involves arithmetic calculation and complex numerical reasoning. * SQA (Pasupat and Liang, 2015) is grounded on documents which contain only a single Table without any surrounding text. The documents are collected from Wikipedia, and the questions are highly compositional, which requires the model with robust table comprehension ability to give correct answers. * HybridDial (Nakamura et al., 2022) is a conversational QA dataset grounded on documents containing both Wikipedia tabular and textual data. The questions are complex which requires reasoning over the documents.
$N=2$ settings - Here we assume a model is available (e.g. via hugging face or pre-trained in-house) that has been pre-trained for autoregressive language modeling on a dataset ($\mathcal{D}_{0}$) using a linear warmup and cosine decay LR schedule. We also assume that the schedule follows existing conventions in the literature (e.g. decaying to the token budget; see Sec. 4 for details) as it is the case for most performant pre-trained LLMs (Rae et al., 2021; Hoffmann et al., 2022; Touvron et al., 2023; 1). Given a model pre-trained on $\mathcal{D}_{0}$, we now assume that a practitioner wants to update this model on a new dataset $\mathcal{D}_{1}$ using the same self-supervised objective.
HIIT differs from other forms of aerobic exercise, such as jogging, cycling, or swimming, in its high-intensity, short-burst nature. While traditional aerobic exercises have a continuous, steady state of intensity, HIIT uses brief periods of intense activity followed by rest, which can result in greater fatigue and thus improved performance.
_Remark C.3_.: Let us conclude by saying that clean data always helps, since $E_{test}$ is decreasing function of $\pi$. Indeed, from (26), the derivative w.r.t $\pi$ is $E^{\prime}_{test}(\pi)=-T\sum_{i\geq k+1}p_{i}^{2}(1-\pi p_{i})^{T-1}\leq 0$.
You will receive an instruction, evidence, and output, and optional preceding sentences. If the preceding sentence is given, the output should be the sentence that follows those preceding sentences. Your task is to evaluate if the output is fully supported by the information provided in the evidence.
Here we present the results for our base models and several other well-known base models across standard academic benchmarks. While benchmarking open-source models, we observed a disparity between the results generated by our pipeline and those reported in public sources. Upon conducting a more in-depth investigation of this difference, mostly because different models use different prompts, post-processing strategies, and sampling techniques. These differences may potentially induce significant variations in the outcomes. Our prompt and post-processing strategy remains consistent with the default settings of the original benchmarks[2; 4; 7; 8; 10; 11; 12; 27; 28; 42; 50; 61; 62; 63; 72; 74; 75; 89; 90]. We use greedy decoding without any post-processing for the generated content. For scores that were not reported publicly (or scores reported with different settings), we try to get results with our pipeline. For scores that can be found publicly, we directly report the existing numbers.
One day, they decided to research the object they had discovered. They scoured the internet, searching for any information about it. It was then that they discovered the true nature of the object - it was an ancient relic from a long-forgotten civilization, imbued with powerful, alien technology.
In this work, we present a novel approach to imbue LLMs with chat capabilities in a new language. Through continual pre-training and the integration of the chat vector into an English-based PLM, the model acquires the ability of instruction following and aligning with human skills, including generating safe responses and multi-turn dialogue. In contrast to the current approach involving CP, SFT, and Human Preference Training, our method entails solely CP and straightforward arithmetic operations, significantly reducing the cost of aligning models with human preferences.
In conversational QA tasks, when a document becomes too lengthy to feed directly into LLMs, a retriever that can handle conversational query becomes essential. This conversational retriever encodes the concatenation of dialogue history and current query, and then retrieve relevant context from documents. After that, only relevant context will be used as inputs for LLMs. The state-of-the-art retrievers, e.g., Dragon (Lin et al., 2023), are optimized for single-turn queries, resulting in a limited generalization capability for multi-turn conversational queries. In Figure 2, we depict our retriever fine-tuning method to alleviate this issue. We propose to use conversational query and context pairs for further fine-tuning a single-turn retriever to better cope with conversational inputs.
\begin{table} \begin{tabular}{l c c c} \hline \hline Model size & 1B & & 7B \\ Dataset & The Pile & RW & RW \\ \hline wiki-bpb $\downarrow$ & 0.64 & 0.66 & 0.60 \\ \hline \hline \end{tabular} \end{table} Table 9: **Models trained on RefinedWeb achieve performance close to models trained on The Pile on Wikitext, despite not having seen any content from Wikipedia.
Improving the quality of retrieval can be considered in two parts: improving the quality of the dataset used for retrieval, and improving the performance of the retrieval technique. Nowadays, many data sets are given to LLM to generate relevant content, and since LLM itself has "hallucination", certain means must be adopted to ensure the accuracy of the data, such as using human beings to supervise the refinement (Chen et al., 2024). In addition, due to the wide range of information sources on the Internet, it is obviously not enough to rely solely on search engines for screening, so it is necessary to improve the retrieval technology, such as the use of BM25 (Luo et al., 2023) or TF-IDF (Lazaridou et al., 2022) algorithms for further re-ranking.
* On Metaculus, for a given question, each prediction of a forecaster is marked by $t$ (starting at 1), from their earliest prediction to the latest. The platform computes the crowd prediction of the question by weighted median.
The experimental results obtained from four code generation benchmarks, namely HumanEval [31], HumanEval+ [32], MBPP [33], and DS-100 [34], demonstrate that our _WizardCoder_ outperforms all other open-source Code LLMs, achieving state-of-the-art (SOTA) performance. Specifically, we observe a substantial improvement in pass@1 scores, with an increase of +22.3 (57.3 vs. 35.0) in HumanEval and +8.2 (51.8 vs. 43.6) in MBPP. Remarkably, despite its much smaller size, our _WizardCoder_ even surpasses Anthropic's Claude and Google's Bard in terms of pass rates on HumanEval and HumanEval+.
Reference [2] Kerala State Film Awards: Kerala Film Critics Award Kerala State Government Award (Professional Drama): Asianet Film Awards: Kerala Kaumadi Gallup Poll Award: Academic G. Venugopal G Venugopal (born 10 December 1960) is an Indian playback singer known for his work in Malayalam films.
This unaddressed uncertainty stems from the challenges inherent in training multiple models of varying sizes and data sizes in traditional scaling experiments. Previously, if the average cost of training one model size on one data size is $C$, then conducting the scaling experiments with $m$ model sizes and $m$ data sizes takes approximately $O(m^{2})C$.
Discussion: It is generally agreed that benchmarking the faithfulness of LLM-generated text is important. However, recent efforts have primarily focused on verifying entity-centric facts (Min et al., 2023). Our work, and others (Zhu et al., 2023; Tang et al., 2024; Mishra et al., 2024), show that these do not provide coverage over all types of LLM errors, especially in more challenging settings like book summarization.
Figure 13: Debate accuracy curves across various datasets using the GPT-3.5-Turbo model.
of other Japanese LLM benchmarks, even exceeding the performance of some previous SOTA 70B parameter Japanese LLMs.
German Benchmarks from (Pluster, 2023), which translated their English counterparts using GPT 3.5 API
We then notice that the PassRatio@256 is significantly lower than that of Pass@256, being 48.2% on GSM8K and 7.9% on MATH. This suggests that while the correct answers to most math questions are present within 256 random generations, there is no assurance that the correct answers will consistently be extracted, a phenomenon we refer to as an "instability issue".
Inspired by Kocetkov et al. (2023); Lee et al. (2021), we removed duplicates and near-duplicates from the training data using Mou et al. (2023), with default parameters, after which we were left with roughly $30$B tokens of high-quality text.
Figure 4: The first step of this problem should be correctly solved as 16 - 3 = 13. We alter the prefix as “16 eggs/day - n eggs/day” to observe the impact of changing 3 to n on the response of SFT and MFT. We give the example of n=1. The orange part is other premises supporting the prediction of the current step.
Furthermore, MEMWALKER Chen et al. (2023) introduces an innovative method to overcome the limitations of context window size in LLMs by creating a memory tree from the input text. This tree is formed by initially segmenting the text into smaller pieces and then summarizing these segments into a hierarchical structure of summary nodes, facilitating efficient indexing and management of large volumes of information.
We employ Llama2-7b-base model for enhancement to eliminate the interference of instruction fine-tuning. We select causal language modeling as our fine-tuning task and create a dataset comprising $200$ documents for each language, extracted from the Wikipedia corpus.4 It is important to note that our enhancements are focused on augmenting the model's capabilities in understanding and generation only; we do not extend its reasoning faculties or broaden its knowledge base as it may require more specific data preparation. Accordingly, we evaluate the efficacy of our enhancements through targeted understanding and generation tasks. Detailed experiment results are shown in Table 8.
$\bullet$_Tensor parallelism._ Tensor parallelism is also a commonly used technique that aims to decompose the LLM for multi-GPU loading. Unlike pipeline parallelism, tensor parallelism focuses on decomposing the tensors (the parameter matrices) of LLMs. For a matrix multiplication operation $Y=XA$ in the LLM, the parameter matrix $A$ can be split into two submatrices, $A_{1}$ and $A_{2}$, by column, which can be expressed as $Y=[XA_{1},XA_{2}]$. By placing matrices $A_{1}$ and $A_{2}$ on different GPUs, the matrix multiplication operation would be invoked at two GPUs in parallel, and the final result can be obtained by combining the outputs from the two GPUs through across-GPU communication. Currently, tensor parallelism has been supported in several open-source libraries, _e.g._, Megatron-LM [75], and can be extended to higher-dimensional tensors. Also, Colossal-AI has implemented tensor parallelism for higher-dimensional tensors [335, 336, 337] and proposed sequence parallelism [338] especially for sequence data, which can further decompose the attention operation of the Transformer model.
Hoffmann et al. (2022) investigate the optimal model size and number of training tokens for training a transformer language model under a given compute budget. The authors train over 400 language models ranging from 70 million to over 16 billion parameters on 5 to 500 billion tokens and find that for compute-optimal training, the model size and number of training tokens should scale at equal rates: for every doubling of model size, the number of training tokens should also be doubled. They then train a model called Chinchilla that is compute-optimal according to their results. For this reason, the scaling laws they propose are often called "Chinchilla scaling laws".
False Positive Analysis. Despite the premises in these experiments being artificially disturbed, we would like to know if MFT might generate false positives due to the reduced use of neighboring features. False positives in this context refer to instances where intermediate steps are incorrect, yet the final result is accurately predicted. This analysis is conducted using GPT-4, which achieves an accuracy of 92% on the specified dataset. The findings, as detailed in Table 6, indicate that MFT does not result in a significant rise in the occurrence of false positives. The model's autoregressive prediction process not only improves the accuracy of the final answer but also preserves the correctness of the intermediate steps. We provide the prompt of GPT-4 and the cases of false positives in Appendix F.
Consider a model $F$ with weight parameters $W\in\mathcal{W}$. Assume $F$ is trained on a $\mathsf{bioD}(N,K,C,D,L,T)$ dataset $\mathcal{Z}$ as defined in Definition 2.2 using any optimizer; this process is represented as $W=W(\mathcal{Z})$ (the model's weight is trained as a function of the training dataset $\mathcal{Z}$). During the evaluation phase, we express $F$ through two functions: $F^{\top}(W,R)$, which generates names, and $F^{\perp}(W,n,a,R)$, which generates values given $(n,a)$, where $R$ denotes the randomness used in generation. Let $F^{\perp}_{1}(W(\mathcal{Z}),n,a,R)$ represent the first chunk of $F^{\perp}(W(\mathcal{Z}),n,a,R)$.
We can alternatively view Theorem 1 as specifying exactly which reward function within each equivalence class the DPO reparameterization selects, that is, the reward function satisfying:
Overall Framework.Next, we first detail the key components of an LLM-based agent and then present the typical workflow.
The fitted curve along the data axis for each dataset and each checkpoints are shown in orange lines in Figure 17.
In Table 4, we further compared our models and OpenAI models across different dataset types in our conversational QA benchmarks. In the comparison between ChatQA-70B and GPT-4, ChatQA-70B achieves better results in text-only documents (avg-text), which demonstrate its superior text understanding ability. While, GPT-4 shows better QA capability in tabular data given the comparisons in avg-table. As for the datasets that require or do not require retrieval, ChatQA-70B and GPT-4 are comparable (gaps are within an average score of 1).
We employed models trained for the Question Answer Generation (QAG) task, specifically using the T5-large model optimized for the End-to-End (E2E) approach, which was reported to deliver superior performance on the QA as the downstream task (Ushio et al., 2023). However, a notable drawback of the E2E method is its limited control over the quality and quantity of generated samples. Additionally, due to the T5 model's input token limitation, the input text must be restricted to 512 tokens. To ensure a fair comparison, we applied the same token limit to the inputs used with the prompt-based approach. Table 4 presents statistics for the QAs generated by both the prompt and E2E methods, using the same corpus.
Stage 3 (new input and output embeddings): At this stage, the input embeddings (embed_tokens) still remain optimized based on the initial embeddings of the output embeddings.
As foundational architectures, we opt for SOLAR-10.7B (Kim et al., 2023) and Phi-2 (Li et al., 2023), because both have shown outstanding performances among similar sizes of LLMs. This choice of foundational architectures aligns with our strategic training objectives, leveraging their proven strengths to ensure our new models achieve similar levels of language understanding and reasoning capabilities in Korean.
* Katherine trained the models used in the paper, built and ran the eval and text generation pipelines, contributed significantly to writing, analysis, and project organization and management. * Daphne ran the approximate matching data deduplication pipelines, extracted prompts and evaluation datasets, ran eval pipelines, and contributed significantly to planning, writing, and analysis. * Andrew wrote the code to perform deduplication with approximate matching, helped evaluate energy expenditure, and helped with analysis. * Chiyuan helped generate plots and contributed to project scoping, writing, and data analysis. * Chris offered mentorship and guidance throughout the project and contributed to writing. * Doug offered mentorship and guidance throughout the project and contributed to writing. * Nicholas wrote the suffix array implementation, ran all ExactSubstr deduplication experiments, contributed significantly to planning, writing, and analysis, as well as scoping the project.
In real applications, there is an increasing demand for long context modeling capacities of LLMs, such as PDF processing and story writing [286]. Many closed-source LLMs provide professional support for long text processing. For instance, OpenAI releases GPT-4 Turbo with a 128K context window, and Anthropic releases Claude 2.1 with a 200K context window. To enhance the long context modeling abilities, there are generally two feasible directions, namely scaling position embeddings and adapting context window. Next, we introduce the two parts in detail.
Figure 2: The Wiki2023 dataset. Top-right: the number of documents and QA pairs; Top-left: frequent keywords in questions; Bottom: the distribution of token counts in documents, questions, and answers.
#### 6.4.2 Final Performance Evaluated by Zero-shot and Few-shot Results on Popular LM Benchmarks
Figure 10: Performance of Baichuan-7B, DeepSeek-7B, DeepSeek-67B, Amber-7B, OpenLLaMA-7B, and Yi-34B models across world knowledge benchmarks.
Despite the extensive knowledge exhibited by LLMs, it is accompanied by several challenges. These encompass the incapacity to access up-to-the-minute news, proclivity towards hallucinations when responding to queries involving out-of-domain knowledge, and the absence of sophisticated reasoning capacities like mathematical calculations or symbolic reasoning. By granting LLMs the ability to employ external tools, it becomes possible to augment the model's reasoning capabilities and assimilate external knowledge, enabling it to engage in information retrieval and environmental interaction.
Since most language tasks can be cast as the prediction problem based on the input, these decoder-only LLMs might be potentially advantageous to implicitly learn how to accomplish these tasks in a unified LM way. Some studies have also revealed that decoder-only LLMs can be naturally transferred to certain tasks by autoregressively predicting the next tokens [26, 55], without fine-tuning. An important variant of LM is the _prefix language modeling_ task, which is designed for pre-training models with the prefix decoder architecture. The tokens within a randomly selected prefix would not be used in computing the loss of prefix language modeling. With the same amount of tokens seen during pre-training, prefix language modeling performs slightly worse than language modeling, since fewer tokens in the sequence are involved for model pre-training [29].
Specifically, current data synthesis approaches fall short of providing effective solutions for task-specific LLM alignment. While prior works by Wang et al. (2022) and Xu et al. (2023) emphasize diversity and complexity as hallmarks of high-quality data, these approaches stumble when facing different downstream tasks that may involve specific instruction distributions. A diverse dataset for one task might not effectively cover the instruction distribution for another. Furthermore, the definition of "complex" instructions can be subjective and vary across tasks. To complicate matters further, an LLM might excel at some seemingly complex instructions while struggling with others that appear simple according to human-crafted criteria. These limitations underscore the need for a unified data synthesis framework that can generate tailored data to align LLMs on specific downstream tasks.
BART(Lewis et al., 2019) is an Encoder-Decoder model that allows for arbitrary noise transformations, as the input to the encoder does not need to align with the output of the decoder. In this case, the document is corrupted by replacing the text span with mask symbols. For pre-training, the researchers proposed five models: Token Masking, Token Deletion, Text Infilling, Sentence Permutation, and Document Rotation. For fine-tuning, the encoder and decoder are fed an uncorrupted document, and the representation of the final hidden state from the decoder is used. Many researchers(Thulke et al., 2021) (Siriwardhana et al., 2023) (Lewis et al., 2019) have adopted BART as the language model in the RALM architecture due to its comprehensive and novel pre-training approach, which greatly enhances the model's robustness.
The Selfmem Cheng et al. (2024) system employs two distinct language models for the machine translation task. The first is a trainable mini-model, which has been trained using a joint and bipartite approach, respectively. The second is a few-shot prompted LLM. Ultimately, Selfmem has demonstrated a notable enhancement in its performance across all four translation directions and for both training architectures. This outcome suggests that enhanced memory capabilities often result in superior generation outcomes. In order to achieve the best results, TRIME Zhong et al. (2022) used the IWSLT'14 De-En baseline. Given that the task is sentence-level, the researchers did not use local memory and long-term memory, as there are few repetitive tokens in them. Instead, they used only external memory, which enabled them to beat the KNN-MT Khandelwal et al. (2020) in performance.
We mainly evaluate WizardMath on two benchmarks (GSM8k [42] and MATH [43]). The GSM8k [42] dataset contains approximately 7500 training data and 1319 test data, mainly on grade school level math problems, each of which consists of basic arithmetic operations (addition, subtraction, multiplication, and division), and generally requires 2 to 8 steps to solve. The MATH [43] dataset collects math problems from prestigious math competitions such as AMC 10, AMC 12, and AIME. It contains 7500 training data and 5,000 challenging test data in seven academic areas: Prealgebra, Algebra, Number Theory, Counting and Probability, Geometry, Intermediate Algebra, and Precalculus. Furthermore, these problems are divided into five levels of difficulty, with '1' denoting the relatively lower difficulty level and '5' indicating the highest level.
The AI community has witnessed the release of several open-source LLMs, where the model weights are made publicly available. EleutherAI has contributed GPT-NeoX-20B [35] and GPT-J-GB [36]. Google has released UL2-20B [37]. Tsinghua University has introduced GLM-130B [7]. Meta has released OPT [9] and LLaMA [8]. It is worth noting that while these open-source models have made valuable contributions, they generally do not exhibit the same level of performance as their closed-source counterparts.
AgentVerse. Multi-agent architectures like AgentVerse demonstrate how distinct phases for group planning can improve an AI agent's reasoning and problem-solving capabilities [2]. AgentVerse contains four primary stages for task execution: recruitment, collaborative decision making, independent action execution, and evaluation. This can be repeated until the overall goal is achieved. By strictly defining each phase, AgentVerse helps guide the set of agents to reason, discuss, and execute more effectively.
Referring to mainstream frameworks such as TransReID [255] and Strong Baseline [264], we retained ID Loss, Triplet Loss, and BN Layer, and replaced the CNN and Transformer backbones using Vim [61] and VAMba [60] to explore the potential of Mamba for re-identification tasks, and the compared results are shown in Table XIII. The selective scanning mechanism (SSM) proposed by the Mamba model allows for sequence modeling with low complexity, and Vim and VAMba further build on it by proposing an SSM modeling approach for 2D image data. Compared to CNN-based models that require complex module design, the simple Mamba network already has effectiveness. Even compared with the models with high complexity such as DeiT [265] and ViT [19], the bidirectional scanning mechanism proposed by Vim has fewer training parameters, and it shows effectiveness on the VehicleID dataset. In contrast, VAMba's cross-scanning mechanism, which does not rely on Transformer's structure (_e.g._, position embedding and class token), has achieved comparable results on the Market1501, DukeMTMC, and VeRi-776 datasets. For this reason, we expect more Mamba-based studies applicable to the re-identification task in the future.
Due to the excellent performance and the long-term decay property, RoPE is widely adopted in the latest LLMs, _e.g._, PaLM [56] and LLMA [57]. Based on RoPE, xPos [277] further improves the translation invariance and length extrapolation of Transformer. At each dimension of the rotation angle vector, xPos adds a special exponential decay that is smaller when the basis is larger. It can alleviate the unstable phenomenon during training as the distance increases.
The main limitation of this article is the relatively small amount of saturated language models we studied. As it is the only suite of language models trained in the range of interest to release an extensive amount of intermediate checkpoints, we could only observe the training dynamics of small Pythia models. Although we observe strong last-layer anisotropy for the smallest GPT-2 model, we cannot tell with certainty whether it suffered from saturation. The OPT-125m model does not display a strong last-layer anisotropy, which could indicate that it was not affected by the saturation phenomenon.
Context LengthAs mentioned, the context length significantly influences the values of the constant terms in scaling-law formulas. Anchoring all constant terms to a specific context length means that we need to rerun the estimation process for every new context length, which is rather inefficient because it is common to adjust the context length to fit various tasks. Given that the loss value at each position also approximately follows a power-law relation [13], it would be possible to include the context length directly as a parameter of the formulas.
K-Adapter(Wang et al., 2021) is another parameter-expansion method that freezes the original parameters of the LM while adding $k$ number of new layers, namely _adapters_, that are updated during continued pretraining. Wang et al. (2021) have shown successful injection of _factual_ and _linguistic_ knowledge for encoder-only models, BERT (Devlin et al., 2019) & RoBERTa (Liu et al., 2019), while we also apply it to an encoder-decoder model, T5, and decoder-only model, GPT-2.
The tool manager in the AIOS system manages a diverse array of API tools that enhance the functionality of LLMs. As shown in Table 1, the tool manager integrates commonly-used tools from various sources [7; 62; 63] and classify them into different categories, which covers web search, scientific computing, database retrieval, image processing, etc. In this way, the managed tools can cover different modalities of input and output (image and text), thus facilitating agent development within the AIOS ecosystem.
An integer sns is odd if sn = 2k+1$ for some integer $k$.n\nTo prove that $A = B$, we must show that $A \subseteq B$ and $B \subseteq A$.
We performed a small experiment to measure our system's performance at recovering entities from prompts. To do this, we created a dataset of prompts that include personal information from Wikibio and Synthbio. We extracted the entities themselves from the tabular portion of the bio datasets and inserted entities into manually-crafted template strings based on Instructions-2M. We release this dataset publicly to aid future research into PII reconstruction.
We train a transformer model with $dim=2048$, $n\_layers=24$, and $n\_heads=16$, resulting in 1.4B parameters. Training data and hyperparameter setup mostly repeats the one from Touvron et al. (2023b). To adapt for the relatively smaller model size, we increase the learning rate to $4.0e-4$, and the global batch-size was capped at 2M due to the limited number of GPUs. During training, we observe a fixed gap in training perplexity between baseline models and reverse training (Figure 1). The loss of the baseline model is measured on data in the standard direction, while the reverse training loss covers data in both directions. We posit that the reverse training doesn't interfere with forward learning -- thus, the model's performance does not degrade on standard benchmarks in data-match conditions, and because we observe a match in the convergence rate of the reverse trained models with the baseline model when it's trained on about 50% of the data.
Answering this question is important.
Figure. 2 (Right) shows the results of models using a mixture of GSM8K and MATH in a 1:1 ratio. These models exhibit balanced scaling behaviors in both in-domain and out-of-domain benchmarks.
$\bullet$The generalization performance of SSMs still deserves attention and further research and improvement. Compared with the limited receptive field and greater complexity of CNN and the Transformer, SSMs have linear complexity and global receptive fields, which may have greater advantages and potential in the field of domain generalization. However, current SSM based networks illustrate limited domain generalization ability, as noted in DGMamba [120]. Long et al. [120] attempt to address this issue from the perspective of hidden states and inappropriate scan mechanisms by proposing the Hidden State Suppressing (HSS) and Semantic-aware Patch Refining (SPR) strategies. We believe more insights and improvements can be conducted to further improve the overall performance of domain generalization.
We intend to continuously support and extend OLMo and its framework, and continue to push the boundaries of open LMs to empower the open research community. To that end, we look forward to bringing different model sizes, modalities, datasets, safety measures, and evaluations into the OLMo family. We hope this and future releases will empower and strengthen the open research community and inspire a new wave of innovation.
Figure A.9: Auto-evaluation rated DDx accuracy on all cases. (a) Top-k auto-evaluation rating of AMIE and PCP with respect to the ground truth. Significant (with FDR correction) for $k>2$. (b) Top-k auto-evaluation rating of AMIE and PCP with respect to the accepted differential. Significant (with FDR correction) for $k>4$.
We now present evaluation results for both our EEVE-Korean and EEVE-Korean-Instruct variants with other top-performing models in Table 3. EEVE-Korean-10.8B-v1.0 outperforms other pre-trained models of similar sizes in the average performance. It is noteworthy that, EEVE-Korean is the only case where the performance in Korean is improved without compromising the performance in English. For example, though OPEN-SOLAR-KO-10.7B, which is built on the same base model as ours, performs slightly better than our EEVE-Korean-Instruct-10.8B-v1.0, it fails to preserve the English capabilities, showing lower performance in English tasks than its base model, SOLAR-10.7B-v1.0. We observe similar trends even for our smaller model, EEVE-Korean-2.8B-v1.0 in comparison with the phi-2-ko-v0.1 model, sharing Phi-2 as its base model. This demonstrates the effectiveness of our training strategy, especially considering that we used even fewer training tokens than our competitors.
Motivated by the intuition that these duplicate-driven clusters need tshould be pruned, as well as the recent success of pruning methods in vision and vision-language models [1, 47], we focus our efforts on data selection strategies that manipulate data points based on their position in an embedding space. We embed each document by feeding it into a 125M OPT model and use the last-layer embedding of the last token (we experiment with different embedding spaces in Section A.7).
Continuous Prompt Optimization. Different from discrete prompts, continuous prompts consist of a set of continuous embeddings, which can be directly optimized through the gradient update based on the loss of downstream tasks. Note that continuous prompt optimization has been mainly studied in PLMs, but draws limited attention in era of LLMs due to their massive magnitudes of parameters. We include the discussion of this part for content completeness. In prior work, most studies typically rely on supervised learning to train continuous prompts based on task data. Furthermore, in data-scarce scenarios, transfer learning methods can be employed to alleviate the lack of labeled data on target tasks. These two approaches are detailed below.
Tail phenomenaSelf-Instruct depends on LMs, and it will inherit all the limitations that carry over with LMs. As recent studies have shown (Razeghi et al., 2022; Kandpal et al., 2022), _tail phenomena_ pose a serious challenge to the success of LMs. In other words, LMs' largest gains correspond to the frequent uses of languages (head of the language use distribution), and there might be minimal gains in the low-frequency contexts. Similarly, in the context of this work, it would not be surprising if the majority of the gains by Self-Instruct are skewed toward tasks or instructions that present more frequently in the pretraining corpus. As a consequence, the approach might show brittleness with respect to uncommon and creative instructions.
Figure 1: Overview of Self-Rag. Self-Rag learns to retrieve, critique, and generate text passages to enhance overall generation quality, factuality, and verifiability.
Recently, some researchers have constructed an intermediate module to coordinate the activities of both the retriever and the language model due to space or black-box LLM constraints, without improving either.
We build the infrastructure supporting the full-stack data processing, pretraining, finetuning, and serving. Our infrastructure features: (1). automated managing and monitoring the computing resource; (2). improved the training speed from optimized parallel strategies, kernel efficiency, and long-context support; (3). unified finetuning framework supporting heterogeneous distributed training backend, such as simultaneously using Megatron and DeepSpeed for multiple models in Direct Preference Optimization (DPO) [59]; (4). reducing the deployment cost by various LLM serving accelerations such as quantization, continuous batching, and paged attention. Below we explain these techniques one by one.
How did we create the taxonomy?We created the taxonomy through an iterative process over 10 weeks. It was primarily created by a small number of the white paper authors, with input from the Working Group attendees at weekly meetings. We also interviewed 12 civil society experts who are not part of the WG, and reviewed 25+ existing taxonomies, 50+ AI safety evaluation datasets, 50+ research and policy papers, and 10+ Community guidelines from industry Trust and Safety orgs. The definitions, subcategories, and sub-subcategories of the taxonomy were refined by writing and reviewing prompts. In the Appendix we offer a side-by-side comparison of our taxonomy against 17 others, and show that our taxonomy contains similar categories to others that have been released (see Appendix A).