In accordance with regulations and requirements, the editorial department's website domain has been changed to arocmag.cn. The original domain (arocmag.com) has been redirecting to new domain since Jan. 1st, 2025.
See More
App. Res. of Comp. Accepted Paper
Vol. 42, 2025 No. 12

Data pruning method based on consistency committee of small-scale heterogeneous language models

Wang Kaiwen
Wang Yunzhe
Tan Wei
Fu Qiming
Lu You
Chen Jianping
School of Electronic & Information Engineering, Suzhou University of Science & Technology, Suzhou Jiangsu 215009, China
DOI: 10.19734/j.issn.1001-3695.2025.05.0139

Abstract

Large language models (LLMs) fine-tuning performance strongly depends on the quality of training data. Existing single-model perplexity-based data evaluation methods had limitations, including perplexity bias, where low-perplexity samples could still be mispredicted, and cross-model divergence, where different models produced inconsistent perplexity scores on the same samples. To address these issues, this study proposed a method based on the consistency of a heterogeneous committee of small language models to evaluate data value from two perspectives. The method calculated the coefficient of variation of perplexity across multiple models to measure model divergence. It also computed prediction difficulty by combining the similarity between predicted outputs and reference answers. Based on these evaluations, the algorithm proposed the MMCS (Multi-Model Consistency Score) metric to select high-quality training data. Experimental results showed that data filtered by MMCS achieved better fine-tuning performance than traditional methods on two mainstream LLMs and three public datasets. Among 36 comparative experiments, 27 obtained optimal results. This finding provides a new approach for efficient data pruning. The evaluation method based on multi-model divergence proves effective in improving the marginal utility of training data.

Foundation Support

国家自然科学基金资助项目(62102278)

Publish Information

DOI: 10.19734/j.issn.1001-3695.2025.05.0139
Publish at: Application Research of Computers Accepted Paper, Vol. 42, 2025 No. 12

Publish History

[2025-07-30] Accepted Paper

Cite This Article

王凯文, 王蕴哲, 谈威, 等. 基于小规模异构语言模型一致性委员会的数据剪枝方法 [J]. 计算机应用研究, 2025, 42 (12). (2025-08-06). https://doi.org/10.19734/j.issn.1001-3695.2025.05.0139. (Wang Kaiwen, Wang Yunzhe, Tan Wei, et al. Data pruning method based on consistency committee of small-scale heterogeneous language models [J]. Application Research of Computers, 2025, 42 (12). (2025-08-06). https://doi.org/10.19734/j.issn.1001-3695.2025.05.0139. )

About the Journal

  • Application Research of Computers Monthly Journal
  • Journal ID ISSN 1001-3695
    CN  51-1196/TP

Application Research of Computers, founded in 1984, is an academic journal of computing technology sponsored by Sichuan Institute of Computer Sciences under the Science and Technology Department of Sichuan Province.

Aiming at the urgently needed cutting-edge technology in this discipline, Application Research of Computers reflects the mainstream technology, hot technology and the latest development trend of computer application research at home and abroad in a timely manner. The main contents of the journal include high-level academic papers in this discipline, the latest scientific research results and major application results. The contents of the columns involve new theories of computer discipline, basic computer theory, algorithm theory research, algorithm design and analysis, blockchain technology, system software and software engineering technology, pattern recognition and artificial intelligence, architecture, advanced computing, parallel processing, database technology, computer network and communication technology, information security technology, computer image graphics and its latest hot application technology.

Application Research of Computers has many high-level readers and authors, and its readers are mainly senior and middle-level researchers and engineers engaged in the field of computer science, as well as teachers and students majoring in computer science and related majors in colleges and universities. Over the years, the total citation frequency and Web download rate of Application Research of Computers have been ranked among the top of similar academic journals in this discipline, and the academic papers published are highly popular among the readers for their novelty, academics, foresight, orientation and practicality.

Indexed & Evaluation

  • The Second National Periodical Award 100 Key Journals
  • Double Effect Journal of China Journal Formation
  • the Core Journal of China (Peking University 2023 Edition)
  • the Core Journal for Science
  • Chinese Science Citation Database (CSCD) Source Journals
  • RCCSE Chinese Core Academic Journals
  • Journal of China Computer Federation
  • 2020-2022 The World Journal Clout Index (WJCI) Report of Scientific and Technological Periodicals
  • Full-text Source Journal of China Science and Technology Periodicals Database
  • Source Journal of China Academic Journals Comprehensive Evaluation Database
  • Source Journals of China Academic Journals (CD-ROM Version), China Journal Network
  • 2017-2019 China Outstanding Academic Journals with International Influence (Natural Science and Engineering Technology)
  • Source Journal of Top Academic Papers (F5000) Program of China's Excellent Science and Technology Journals
  • Source Journal of China Engineering Technology Electronic Information Network and Electronic Technology Literature Database
  • Source Journal of British Science Digest (INSPEC)
  • Japan Science and Technology Agency (JST) Source Journal
  • Russian Journal of Abstracts (AJ, VINITI) Source Journals
  • Full-text Journal of EBSCO, USA
  • Cambridge Scientific Abstracts (Natural Sciences) (CSA(NS)) core journals
  • Poland Copernicus Index (IC)
  • Ulrichsweb (USA)