Pith. sign in

REVIEW 4 major objections 7 minor 1 cited by

Building a Family of Data Augmentation Models for Low-cost LLM Fine-tuning on the Cloud

T0 review · 4 major / 7 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Small open-weight models can generate the data that expensive LLM APIs provide for fine-tuning, at a fraction of the cost.

desk verdict The system and released models are useful, but the headline fine-tuning gains are not actually demonstrated because Table 4 compares against an unfine-tuned baseline, not against fine-tuning on the seed data alone. read the letter →

arxiv 2412.04871 v1 pith:VZYJ2H6I submitted 2024-12-06 cs.CL

Yuanhao Yue , Chengyu Wang , Jun Huang , Peng Wang This is my paper · ORCID
classification cs.CL
keywords dataaugmentationinstructiontuningLLMfine-tuningexpansionrefinementresponsegenerationknowledgedistillationcloudplatform
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that the costly step of collecting and annotating domain-specific data for LLM fine-tuning can be handed to small, specialized models. It builds three cheap assistants—instruction expander, instruction refiner, and instruction-response pair expander—trained on 36K seed pairs distilled from larger LLMs. On BIG-Bench tasks, fine-tuning with expanded data adds 1-2 percentage points on Elementary Math and 7-11 points on Implicature. A user study finds the 7B refiner matches or beats a much larger proprietary API on prompt refinement. If these gains hold broadly, domain fine-tuning becomes dramatically cheaper.

What carries the argument

The automatic data collection system is the engine: a Data Source Collector assembles a 36K-pair seed set from public repositories, in-house samples, and machine translation, balanced by task-aware sampling; an LLM-Based Augmenter (Qwen-max for Chinese, GPT-4 otherwise) generates expansion, refinement, and response-generation targets with quality filtering; a Training Set Generator organizes these into three datasets ($D_{IE}$, $D_{IR}$, $D_{IRE}$). The student models are trained with three autoregressive losses: $\mathcal{L}_{IE}$ for expanding one source instruction into several same-task instructions, $\mathcal{L}_{IR}$ for refining an instruction, and $\mathcal{L}_{IRE}$ for producing new instruction-response pairs given $K$ in-context examples. This distillation chain compresses the augmentation ability of the large teachers into small, cheap models.

What would settle it

Take a deliberately narrow or biased seed set (e.g., only one task type or one topic domain), run the full augmentation and fine-tuning pipeline, and compare against fine-tuning on the un-augmented seed. If model performance does not improve over the seed-only baseline, the claimed augmentation benefit is an artifact of the seed's coverage. A second check: rebuild the seed from a different public corpus and see whether the 1-2 point Elementary Math and 7-11 point Implicature gains reproduce.

Watch

Extended reading notes

Core claim

The central claim is that small open-weight chat models (Qwen2-1.5B and 7B) can be fine-tuned to perform the three data augmentation functions—instruction expansion, instruction refinement, and instruction-response pair expansion—at a fraction of the inference cost of closed-source LLM APIs, and that the data they generate improves downstream fine-tuning as effectively as data produced by the proprietary teachers. The paper establishes this by training each model on augmentation tuples generated by Qwen-max and GPT-4, then measuring (i) fine-tuning gains on Elementary Math and Implicature, (ii) relative win rates for response detail and truthfulness after prompt refinement on MT-Bench, and (iii) diversity, length, complexity, and factuality of instruction-response expansion against Self-Instruct. The strongest quantitative evidence is that fine-tuning on expanded data consistently beats fine-tuning on seed data alone, with the largest gains on tasks the base model was not specialized in.

Load-bearing premise

The whole pipeline stands on the quality and diversity of the 36K seed instruction-response pairs: if the seed is biased or narrow, the augmentation models inherit and amplify those flaws, and the fine-tuning gains may vanish.

Editorial extensions

If this is right

  • Users can fine-tune LLMs on domains without paying for large closed-source APIs to build datasets; small augmentation models do the expansion at low inference cost.
  • Fine-tuning on augmented data consistently outperforms seed-only fine-tuning, with larger gains on tasks the base model has not been specialized in (Implicature: +7-11 points).
  • The 7B refinement model can replace a several-hundred-billion-parameter proprietary model for prompt refinement, matching or beating it in win-lose-tie rates.
  • Instruction-response pair expansion produces more diverse and complex data than Self-Instruct at similar factuality, which should improve downstream training.
  • Because the augmentation models are small, the approach can be embedded into a cloud ML platform as a low-cost data-preparation service.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The seed-set bottleneck suggests a natural stress test: augmenting from a seed that deliberately under-represents a target domain should fail, meaning practitioners must audit seed coverage before relying on the pipeline.
  • The same distillation chain could be re-run for other languages beyond English and Chinese by swapping the machine-translation step, or for specialized modalities (code, tables) provided seed pairs exist.
  • The iterative design hints at a self-improving loop: data produced by the trained augmentation models could be filtered and folded back into the seed to bootstrap further gains, though the paper does not test this.
  • Cost comparisons in the paper focus on inference; a full cloud accounting would also need to include the one-time distillation cost of building the seed and training the small models.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The paper presents a family of small (1.5B/7B) Qwen2-based data augmentation models for three operations: instruction expansion, instruction refinement, and instruction-response pair expansion. Training data are produced by an automatic collection system that starts from 36K seed instruction-response pairs gathered from public repositories, in-house data, and machine translation, then uses GPT-4 and Qwen-max to expand, refine, and rewrite instructions and responses with quality checks. The models are trained with the three corresponding causal language-modeling losses, and the resulting checkpoints are released and integrated into a cloud ML platform. Experiments report fine-tuning gains on Elementary Math and Implicature (Table 4), refinement win rates on MT-Bench judged by GPT-4 (Table 5), diversity/complexity/factuality comparisons against Self-Instruct (Table 6), and a prompt-refinement application study versus Qwen-max (Figure 3). The paper concludes that the approach improves the efficiency of LLM fine-tuning and reduces cost.

Significance. The practical motivation is clear: small, cheap augmentation models that can replace expensive closed-source APIs for dataset construction would be useful to practitioners, and the paper is commendable for releasing the five checkpoints and describing a deployed platform integration. If the fine-tuning gains in Table 4 were established with proper controls, the contribution would be valuable. However, as presented, the central quantitative claims are not yet supported by the experimental design: the main instruction-expansion experiment lacks control arms, the refinement and application evaluations rely entirely on GPT-4 judgments without human validation, and the instruction-response expansion comparison is against a single baseline. The paper therefore reads more as a system/technical report than as a fully validated scientific study; the released models and pipeline description are the strongest assets.

major comments (4)
  1. [§4.1, Table 4] The attribution of the reported 1-2 percentage point and 7-11 percentage point gains to the instruction-expansion models is not established by the current protocol. The expanded set is generated by the IE model, but responses are annotated by Qwen-max, and the comparison is between an off-the-shelf Qwen2-Instruct model and a model fine-tuned on this expanded, Qwen-max-annotated data. Any supervised fine-tuning on additional high-quality instruction-response pairs, regardless of how the instructions were produced, could produce the gains. The paper should add control arms: fine-tuning on the original seed only, fine-tuning on an equally sized seed expanded by a non-learned method such as Self-Instruct or paraphrase with the same Qwen-max response annotation, and fine-tuning on duplicated seed examples to control for dataset size. Without such arms, the numbers in Table 4 cannot support the claim in §4.1 that the data augmentation technique 'can still consistently improve the model's performance.'
  2. [§4.2, Table 5; §4.4, Figure 3] Both the MT-Bench refinement evaluation and the application study use GPT-4-turbo as the sole judge, with no human evaluation, inter-annotator agreement, or validation of the judge reported. Since the IR and IE models were trained partly from GPT-4 outputs, using the same model family as judge creates a non-independence risk, and for a paper whose headline claim is improved response quality, an automated judge without validation is insufficient. At minimum, the authors should report a human-validated subset with agreement statistics or use an independently verified judge.
  3. [§4.3, Table 6] The instruction-response expansion evaluation compares only against Self-Instruct and uses proxy metrics: unique bigrams for diversity, average token count for length, average IFD perplexity for complexity, and GPT-4 factuality scores. These metrics have no reported variance, and no downstream task measures whether the augmented data improves fine-tuning. The claim in Table 6 that truthfulness 'approaches that of Self-Instruct' rests on a 0.1 difference in a 1-5 scale. A task-based evaluation, such as using data generated by each method to fine-tune the same model and measuring performance on held-out tasks, would be needed to support the effectiveness claim.
  4. [§3.1.1, Limitations] The automatic data collection system is the sole source of training data for all augmentation models, and the paper's own Limitations section acknowledges that performance is tied to seed quality and diversity. However, no analysis of the seed composition is given, such as task distribution, language balance, noise level, or overlap with evaluation sets, and no leakage check is reported for the BIG-Bench subsets used in §4.1. Since the system relies on in-house data that is not described in sufficient detail for reproduction, this is a substantive reproducibility gap. A minimal addition would be a breakdown of the 36K pairs by source and task type, plus a leakage analysis against the evaluation benchmarks.
minor comments (7)
  1. [§4.1] The phrase 'we split a subset of 100 data instances as seed dataset' is ambiguous; the paper should report how the remaining instances were used for evaluation and give the evaluation set sizes for both BIG-Bench tasks.
  2. [Table 4] For Implicature, Qwen2-7B with Qwen2-7B-Instruct-Exp (32.92%) is lower than Qwen2-7B with Qwen2-1.5B-Instruct-Exp (35.41%); this non-monotonicity is not discussed, though it is relevant to the claim that larger expansion models are consistently better.
  3. [Table 6] No error bars or multiple runs are reported for any of the four metrics, and a 0.1 difference in factuality (5.0 vs 4.9) is treated as evidence of comparable performance; this needs statistical support.
  4. [§4.4] The term 'user study' appears to describe an automated comparison with GPT-4-turbo as judge rather than a study with human participants; please use a less misleading term or clarify the design.
  5. [Throughout] Minor typos and inconsistencies include 'augementation' in footnote 1, 'sever' for 'server' in §4, and inconsistent hyphenation of 'task-aware sampling'.
  6. [References] The two Li et al. 2023 and 2024 references appear to be the same work (arXiv:2308.12032) cited twice; please consolidate them.
  7. [Figure 3] The win-lose-tie rates are presented without the number of samples or the exact percentages; these should be reported in the text.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the paper's quantitative claims are evaluated on external benchmarks, and no fitted parameter or self-citation reduces the predicted gains to the model's training inputs by construction.

full rationale

The derivation chain in this paper is empirical rather than algebraic. The augmentation models are trained with the losses in Eqs. (1)-(3) on datasets D_IE, D_IR, and D_IRE, whose contents are generated by Qwen-max and GPT-4 from a 36K seed collection. The paper's central claims are then tested on independent instrumentation: Table 4 fine-tunes Qwen2-Instruct models on BIG-Bench seeds expanded by the IE models and reports downstream accuracy; Table 5 measures win rates on MT-Bench; Table 6 compares IRE output against Self-Instruct on diversity, length, complexity, and factuality; and Figure 3 compares the 7B refine model against Qwen-max using GPT-4-turbo as judge. None of these quantities is defined in terms of the augmentation model's own training objective, and there is no equation in which the reported gain equals the fit or the generating distribution. The self-citations to Yue et al. (2024) appear only to name task-aware sampling and a general distillation idea; they are not used as a uniqueness theorem or as the justification for the headline improvements. The limitations acknowledged by the authors, including dependence on seed quality and public LLMs, are external dependencies and evaluation risks (e.g., GPT-4 judging outputs from a pipeline that used GPT-4 as a teacher, and the absence of a same-data-size control in Table 4), not definitional circularity. Therefore no circular step can be exhibited with the required specificity.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

No free parameters are fitted to data in the sense of adjusting constants to force a result. The training hyperparameters (lr=1e-5, 3 epochs) are standard choices. The axioms are domain assumptions about data quality, teacher reliability, distillation capacity, and evaluation metrics, all of which are load-bearing for the central claim.

assumptions (4)
  • domain assumption The 36K-pair seed dataset assembled from public repositories, in-house data, and machine translation is diverse, task-balanced, and high-quality enough to train augmentation models that generalize.
    Section 3.1.1 states that task-aware sampling is used to balance the distribution, but no quantitative diversity or quality metrics on the seed set are given. The paper's Limitations section acknowledges the dependency on seed data quality.
  • domain assumption The proprietary LLMs (Qwen-max and GPT-4) produce factually correct and useful augmented instructions and responses, and the LLM-based quality checker reliably filters out low-quality outputs.
    Section 3.1.2 describes the LLM-based augmenter and quality check, but provides no human evaluation of the generated data or the checker's precision. The entire training set is synthetic.
  • domain assumption Distillation from large proprietary LLMs into 1.5B and 7B models preserves enough augmentation capability for the target tasks.
    Section 3.2 trains the small models on teacher-generated data. The paper assumes the small models can learn the augmentation functions, but offers no analysis of capacity limits, and the 1.5B model is not used for response expansion because it 'lacks capacity'.
  • domain assumption The evaluation metrics used for instruction-response expansion (unique bigrams, IFD complexity, perplexity, and GPT-4 factuality) are valid proxies for data quality in fine-tuning.
    Section 4.3 compares IRE output to Self-Instruct on these metrics, but does not show that higher diversity/length/complexity leads to better downstream fine-tuning, which is the practical claim of the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Building a Family of Data Augmentation Models for Low-cost LLM Fine-tuning on the Cloud." pith.science (2026). https://pith.science/paper/VZYJ2H6I

@misc{pith2026241204871,
  author       = {Pith},
  title        = {Pith review of: Building a Family of Data Augmentation Models for Low-cost LLM Fine-tuning on the Cloud},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VZYJ2H6I}},
  note         = {Machine review of arXiv:2412.04871}
}
read the original abstract

Specializing LLMs in various domain-specific tasks has emerged as a critical step towards achieving high performance. However, the construction and annotation of datasets in specific domains are always very costly. Apart from using superior and expensive closed-source LLM APIs to construct datasets, some open-source models have become strong enough to handle dataset construction in many scenarios. Thus, we present a family of data augmentation models designed to significantly improve the efficiency for model fine-tuning. These models, trained based on sufficiently small LLMs, support key functionalities with low inference costs: instruction expansion, instruction refinement, and instruction-response pair expansion. To fulfill this goal, we first construct an automatic data collection system with seed datasets generated from both public repositories and our in-house datasets. This system leverages powerful LLMs to expand, refine and re-write the instructions and responses, incorporating quality assessment techniques. Following this, we introduce the training process of our models, which effectively distills task-solving and text synthesis abilities from teacher LLMs. Finally, we demonstrate how we integrate these functionalities into a machine learning platform to support low-cost LLM fine-tuning from both dataset preparation and training perspectives for users. Experiments and an application study prove the effectiveness of our approach.

Figures

Figures reproduced from arXiv: 2412.04871 by the authors.

Figure 1
Figure 1. The data collection system. our trained models exhibits versatility and can be deployed across a diverse range of NLP tasks based on the instruction tuning paradigm. 3 The Proposed Approach In this section, we present our work on data aug￾mentation models for low-cost LLM fine-tuning. 3.1 Data Collection System The high-level architecture of our data collection system is shown in [PITH_FULL_IMAGE:figures/full_fig_p… view at source ↗
Figure 2
Figure 2. A snapshot of the model card. Model Math Impl. Qwen2-1.5B-Instruct 57.90% 28.96% + Qwen2-1.5B-Instruct-Exp 59.15% 31.22% + Qwen2-7B-Instruct-Exp 58.32% 39.37% Qwen2-7B-Instruct 71.40% 28.85% + Qwen2-1.5B-Instruct-Exp 73.90% 35.41% + Qwen2-7B-Instruct-Exp 72.53% 32.92% [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The win-lose-tie rates of Qwen2-7B-Instruct￾Refine for the prompt refinement task, compared with the much larger model Qwen-max. their effective use can be challenging for begin￾ners without experiences to craft detailed and infor￾mative prompts. Therefore, LLMs are commonly employed as prompt engineers to enhance user experience. In a mobile chatbot application, the chat pipeline integrates a large proprietary LLM,… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: We observe that the data generated by Qwen2-7B-Instruct-Response-Exp, compared to data generated by Self-Instruct, occupies a more broadly distributed range of regions within the embedding space after being projected to two dimensions using t-SNE. 60 40 20 0 20 40 60 8…
Figure 5
Figure 5. Figure 5: Distribution of the model expansion and human-written dataset in the embedding space on the Elementary [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. EasyDistill: A Comprehensive Toolkit for Effective Knowledge Distillation of Large Language Models

    cs.CL 2025-05 conditional novelty 6.0 of 10

    EasyDistill packages established LLM knowledge-distillation techniques into a single modular toolkit with released distilled models, datasets, and Alibaba Cloud integration.

Reference graph

Works this paper leans on

38 extracted references · 5 canonical work pages · cited by 1 Pith paper

  1. [1]

    Amirhossein Abaskohi, Sascha Rothe, and Yadollah Yaghoobzadeh. 2023. https://doi.org/10.18653/V1/2023.ACL-SHORT.59 LM-CPPF: paraphrasing-guided data augmentation for contrastive prompt-based few-shot fine-tuning . In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), ACL 2023, Toronto, Canada,...

    doi:10.18653/v1/2023.acl-short.59 2023

  2. [2]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng X...

    arXiv doi:10.48550/arxiv.2309.16609 2023

  3. [3]

    BIG bench authors. 2023. https://openreview.net/forum?id=uyTL5Bvosj Beyond the imitation game: Quantifying and extrapolating the capabilities of language models . Transactions on Machine Learning Research

    2023

  4. [4]

    Yu, Qiang Yang, and Xing Xie

    Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, Wei Ye, Yue Zhang, Yi Chang, Philip S. Yu, Qiang Yang, and Xing Xie. 2024. https://doi.org/10.1145/3641289 A survey on evaluation of large language models . ACM Trans. Intell. Syst. Technol. , 15(3):39:1--39:45

    doi:10.1145/3641289 2024

  5. [5]

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. http://papers.nips.cc/paper\_files/paper/2023/hash/1feb87871436031bdc0f2beaa62a049b-Abstract-Conference.html Qlora: Efficient finetuning of quantized llms . In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, Neur...

    2023

  6. [6]

    Jacob Devlin, Ming - Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://doi.org/10.18653/V1/N19-1423 BERT: pre-training of deep bidirectional transformers for language understanding . In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, ...

    doi:10.18653/v1/n19-1423 2019

  7. [7]

    Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. 2022. https://doi.org/10.18653/V1/2022.ACL-LONG.26 GLM: general language model pretraining with autoregressive blank infilling . In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2022, Dublin, Irelan...

    doi:10.18653/v1/2022.acl-long.26 2022

  8. [8]

    Feng, Varun Gangal, Jason Wei, Sarath Chandar, Soroush Vosoughi, Teruko Mitamura, and Eduard H

    Steven Y. Feng, Varun Gangal, Jason Wei, Sarath Chandar, Soroush Vosoughi, Teruko Mitamura, and Eduard H. Hovy. 2021. https://doi.org/10.18653/V1/2021.FINDINGS-ACL.84 A survey of data augmentation approaches for NLP . In Findings of the Association for Computational Linguistics: ACL/IJCNLP 2021, Online Event, August 1-6, 2021 , volume ACL/IJCNLP 2021 of F...

    doi:10.18653/v1/2021.findings-acl.84 2021

Show all 38 references
  1. [9]

    Cheng - Yu Hsieh, Chun - Liang Li, Chih - Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alex Ratner, Ranjay Krishna, Chen - Yu Lee, and Tomas Pfister. 2023. https://doi.org/10.18653/V1/2023.FINDINGS-ACL.507 Distilling step-by-step! outperforming larger language models with less tr...

    2023 doi

  2. [10]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen - Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen - Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. https://openreview.net/forum?id=nZeVKeeFYf9 Lora: Low-rank adaptation of large language models . In The Tenth International Conference on Learning Representat...

    2022

  3. [11]

    Dibyakanti Kumar, Vivek Gupta, Soumya Sharma, and Shuo Zhang. 2022. https://doi.org/10.18653/V1/2022.FINDINGS-EMNLP.324 Realistic data augmentation framework for enhancing tabular reasoning . In Findings of the Association for Computational Linguistics: EMNLP 2022, Abu Dhabi, ...

    2022 doi

  4. [12]

    Ming Li, Yong Zhang, Zhitao Li, Jiuhai Chen, Lichang Chen, Ning Cheng, Jianzong Wang, Tianyi Zhou, and Jing Xiao. 2023. https://doi.org/10.48550/ARXIV.2308.12032 From quantity to quality: Boosting LLM performance with self-guided data selection for instruction tuning . CoRR, a...

    2023 doi arXiv

  5. [13]

    Ming Li, Yong Zhang, Zhitao Li, Jiuhai Chen, Lichang Chen, Ning Cheng, Jianzong Wang, Tianyi Zhou, and Jing Xiao. 2024. https://arxiv.org/abs/2308.12032 From quantity to quality: Boosting llm performance with self-guided data selection for instruction tuning . Preprint, arXiv:...

    2024 arXiv

  6. [14]

    Xiao Liu, Xuanyu Lei, Shengyuan Wang, Yue Huang, Zhuoer Feng, Bosi Wen, Jiale Cheng, Pei Ke, Yifan Xu, Weng Lam Tam, Xiaohan Zhang, Lichao Sun, Hongning Wang, Jing Zhang, Minlie Huang, Yuxiao Dong, and Jie Tang. 2023. https://doi.org/10.48550/ARXIV.2311.18743 Alignbench: Bench...

    2023 doi arXiv

  7. [15]

    Bonan Min, Hayley Ross, Elior Sulem, Amir Pouran Ben Veyseh, Thien Huu Nguyen, Oscar Sainz, Eneko Agirre, Ilana Heintz, and Dan Roth. 2024. https://doi.org/10.1145/3605943 Recent advances in natural language processing via large pre-trained language models: A survey . ACM Comp...

    2024 doi

  8. [16]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan Leik...

    2022

  9. [17]

    Xipeng Qiu, Tianxiang Sun, Yige Xu, Yunfan Shao, Ning Dai, and Xuanjing Huang. 2020. https://arxiv.org/abs/2003.08271 Pre-trained models for natural language processing: A survey . CoRR, abs/2003.08271

    2020 arXiv

  10. [18]

    Manning, Stefano Ermon, and Chelsea Finn

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D. Manning, Stefano Ermon, and Chelsea Finn. 2023. http://papers.nips.cc/paper\_files/paper/2023/hash/a85b405ed65c6477a4fe8302b5e06ce7-Abstract-Conference.html Direct preference optimization: Your language model is sec...

    2023

  11. [19]

    Gaurav Sahu, Olga Vechtomova, Dzmitry Bahdanau, and Issam H. Laradji. 2023. https://doi.org/10.18653/V1/2023.EMNLP-MAIN.323 Promptmix: A class boundary augmentation method for large language model distillation . In Proceedings of the 2023 Conference on Empirical Methods in Nat...

    2023 doi

  12. [20]

    Kashun Shum, Shizhe Diao, and Tong Zhang. 2023. https://doi.org/10.18653/V1/2023.FINDINGS-EMNLP.811 Automatic prompt augmentation and selection with chain-of-thought from labeled data . In Findings of the Association for Computational Linguistics: EMNLP 2023, Singapore, Decemb...

    2023 doi

  13. [21]

    Ross Taylor, Marcin Kardas, Guillem Cucurull, Thomas Scialom, Anthony Hartshorn, Elvis Saravia, Andrew Poulton, Viktor Kerkez, and Robert Stojnic. 2022. https://doi.org/10.48550/ARXIV.2211.09085 Galactica: A large language model for science . CoRR, abs/2211.09085

    2022 doi arXiv

  14. [22]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie - Anne Lachaux, Timoth \' e e Lacroix, Baptiste Rozi \` e re, Naman Goyal, Eric Hambro, Faisal Azhar, Aur \' e lien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023 a . https://doi.org/10....

    2023 doi arXiv

  15. [23]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton - Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu,...

    2023 doi arXiv

  16. [24]

    Yida Wang, Pei Ke, Yinhe Zheng, Kaili Huang, Yong Jiang, Xiaoyan Zhu, and Minlie Huang. 2020. https://doi.org/10.1007/978-3-030-60450-9\_8 A large-scale chinese short-text conversation dataset . In Natural Language Processing and Chinese Computing - 9th CCF International Confe...

    2020 doi

  17. [25]

    Smith, Daniel Khashabi, and Hannaneh Hajishirzi

    Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. https://doi.org/10.18653/v1/2023.acl-long.754 Self-instruct: Aligning language models with self-generated instructions . In Proceedings of the 61st Annual Mee...

    2023 doi

  18. [26]

    Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M

    Jason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V. Le. 2022. https://openreview.net/forum?id=gEZrGCozdqR Finetuned language models are zero-shot learners . In The Tenth International Conference on Learning Repr...

    2022

  19. [27]

    Jules White, Quchen Fu, Sam Hays, Michael Sandborn, Carlos Olea, Henry Gilbert, Ashraf Elnashar, Jesse Spencer - Smith, and Douglas C. Schmidt. 2023. https://doi.org/10.48550/ARXIV.2302.11382 A prompt pattern catalog to enhance prompt engineering with chatgpt . CoRR, abs/2302.11382

    2023 doi arXiv

  20. [28]

    Dingjun Wu, Jing Zhang, and Xinmei Huang. 2023. https://doi.org/10.18653/V1/2023.FINDINGS-ACL.408 Chain of thought prompting elicits knowledge augmentation . In Findings of the Association for Computational Linguistics: ACL 2023, Toronto, Canada, July 9-14, 2023 , pages 6519--...

    2023 doi

  21. [29]

    Yuanhao Yue, Chengyu Wang, Jun Huang, and Peng Wang. 2024. https://doi.org/10.48550/ARXIV.2405.13448 Distilling instruction-following abilities of large language models with task-aware curriculum planning . CoRR, abs/2405.13448

    2024 doi arXiv

  22. [30]

    Shengyu Zhang, Linfeng Dong, Xiaoya Li, Sen Zhang, Xiaofei Sun, Shuhe Wang, Jiwei Li, Runyi Hu, Tianwei Zhang, Fei Wu, and Guoyin Wang. 2023 a . https://doi.org/10.48550/ARXIV.2308.10792 Instruction tuning for large language models: A survey . CoRR, abs/2308.10792

    2023 doi

  23. [31]

    Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer

    Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona T. Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer...

    2022 doi arXiv

  24. [32]

    Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola. 2023 b . https://openreview.net/pdf?id=5NTt8GFjUHkr Automatic chain of thought prompting in large language models . In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 20...

    2023

  25. [33]

    Xing, Hao Zhang, Joseph E

    Lianmin Zheng, Wei - Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. http://papers.nips.cc/paper\_files/paper/2023/hash/91f18a1287b398d378ef22505bf41832-Abstr...

    2023

  26. [34]

    Chunting Zhou, Pengfei Liu, Puxin Xu, Srinivasan Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, Susan Zhang, Gargi Ghosh, Mike Lewis, Luke Zettlemoyer, and Omer Levy. 2023 a . http://papers.nips.cc/paper\_files/paper/2023/hash/ac662d74829e4407ce1d126477f4...

    2023

  27. [35]

    Jing Zhou, Yanan Zheng, Jie Tang, Li Jian, and Zhilin Yang. 2022. https://doi.org/10.18653/V1/2022.ACL-LONG.592 Flipda: Effective and robust data augmentation for few-shot learning . In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Vo...

    2022 doi

  28. [36]

    Meng Zhou, Xin Li, Yue Jiang, and Lidong Bing. 2023 b . https://doi.org/10.18653/V1/2023.FINDINGS-ACL.700 Enhancing cross-lingual prompting with dual prompt augmentation . In Findings of the Association for Computational Linguistics: ACL 2023, Toronto, Canada, July 9-14, 2023 ...

    2023 doi

  29. [37]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  30. [38]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.