面向审计大模型微调的不平衡指令筛选策略

doi:10.3778/j.issn.1673-9418.2507017

摘要/Abstract

摘要： 在面向监管类（如法律咨询、审计判断）的垂直领域大语言模型指令微调中，多任务指令微调数据集中存在高低资源任务指令数据不平衡的问题。现有的指令筛选策略往往忽视了任务之间的协同效应和领域特定要求。为此，提出一种分阶段不平衡指令筛选策略（IIFS），用于从不平衡的多任务指令微调数据集中选择高质量的指令子集。IIFS通过指令冗余性、微调必要性和任务相关性来系统地评估高资源任务指令数据。通过动态聚类过滤冗余数据。使用文本相似度评估指令数据在目标模型上的响应效果以评估该指令微调的必要性。量化高资源任务指令数据与核心任务的语义关系衡量任务相关性。综合质量分与聚类规模构建一种自适应动态采样机制获取一批高质量指令微调数据子集以进行模型微调。基于IIFS采样策略将高低资源任务指令数据比例从11.5∶1降至2.8∶1，冗余率降低了75.61%，在保持领域适应性的同时缓解了数据不平衡性。在评估数据集上的实验表明，基于IIFS微调的大模型比基于完整指令集微调的大模型整体性能提升3.57个百分点；特别的，在审计案例分类任务上F1-score较完整指令集提升4.84个百分点。这项工作为垂直领域的工业规模大模型指令微调提供了一种经济高效的自动化解决方案。

关键词: 大语言模型, 多任务指令微调, 不平衡指令集, 指令筛选策略

Abstract: In the instruction fine-tuning of vertical large language models (LLMs) for regulatory applications (such as legal consultation and audit judgment), multi-task instruction fine-tuning of large language models faces the challenge of imbalanced instruction data between high-resource and low-resource tasks within multi-task instruction datasets. Existing instruction filtering strategies often neglect task synergies and domain-specific requirements. To address these problems, this paper proposes a phased imbalanced instruction filtering strategy (IIFS) for selecting a high-quality instruction subset from an imbalanced instruction dataset for model fine-tuning. IIFS systematically evaluates high-resource task instruction data based on instruction redundancy, fine-tuning necessity, and task relevance. Firstly, it filters redundant data via dynamic clustering. Secondly, the response from the target model to the instruction data is evaluated using text similarity to determine the necessity of this instruction fine-tuning. Thirdly, it quantifies the semantic relationship between high-resource task instructions and core tasks to measure task relevance. Finally, an adaptive dynamic sampling mechanism, integrating quality scores and cluster size, selects a high-quality instruction fine-tuning data subset for fine-tuning. Leveraging the IIFS sampling strategy, the high-to-low-resource task data ratio is reduced from 11.5∶1 to 2.8∶1, with redundancy rate decreased by 75.61%. This effectively mitigates data imbalance while preserving domain adaptability. Experiments on evaluation datasets demonstrate that LLMs fine-tuned by using the IIFS-selected subset achieve an overall performance improvement of 3.57 percentage points compared with models fine-tuned on the full instruction set. Notably, an increase of 4.84 percentage points in F1-score is observed for the audit case classification task. This work provides an economical and efficient automated solution for industrial-scale LLM instruction fine-tuning in vertical domains.

Key words: large language models, multi-task instruction fine-tuning, imbalanced instruction dataset, instruction filtering strategy

黄佳佳, 朱浩然, 姜茂伟, 陈勇, 徐超. 面向审计大模型微调的不平衡指令筛选策略[J]. 计算机科学与探索, 2025, 19(12): 3353-3367.

HUANG Jiajia, ZHU Haoran, JIANG Maowei, CHEN Yong, XU Chao. Imbalanced Instruction Filtering Strategy for Fine-Tuning Audit Large Language Models[J]. Journal of Frontiers of Computer Science and Technology, 2025, 19(12): 3353-3367.

参考文献

[1] OPENAI. GPT-4 technical report[R/OL]. [2025-05-14]. https:// arxiv.org/abs/2303.08774.
[2] DEEPSEEK-AI. DeepSeek-V3 technical report[R/OL]. [2025-05-14]. https://arxiv.org/abs/2412.19437.
[3] QWEN. Qwen2.5 technical report[R/OL]. [2025-05-14]. https://arxiv.org/abs/2412.15115.
[4] CUI J X, NING M N, LI Z J, et al. Chatlaw: a multi-agent collaborative legal assistant with knowledge graph enhanced mixture-of-experts large language model[EB/OL]. [2025-05-14]. https://arxiv.org/abs/2306.16092.
[5] HUANG J J, ZHU H R, XU C, et al. AuditWen: an open-source large language model for audit[C]//Proceedings of the 23rd Chinese National Conference on Computational Linguistics. Singapore: Springer, 2024: 505-521.
[6] ZHANG H B, CHEN J Y, JIANG F, et al. HuatuoGPT, towards taming language model to be a doctor[C]//Findings of the Association for Computational Linguistics: EMNLP 2023. Stroudsburg: ACL, 2023: 10859-10885.
[7] XIE Q Q, HAN W G, ZHANG X, et al. PIXIU: a large language model, instruction data and evaluation benchmark for finance[EB/OL]. [2025-05-14]. https://arxiv.org/abs/2306. 05443.
[8] ZHOU C T, LIU P F, XU P X, et al. LIMA: less is more for alignment[C]//Advances in Neural Information Processing Systems 36, 2023.
[9] CHEN L C, LI S Y, YAN J, et al. AlpaGasus: training a better Alpaca with fewer data[C]//Proceedings of the 12th International Conference on Learning Representations, 2024.
[10] CAO Y H, KANG Y B, WANG C, et al. Instruction mining: instruction data selection for tuning large language models[EB/OL]. [2025-05-14]. https://arxiv.org/abs/2307.06290.
[11] LIU W, ZENG W H, HE K Q, et al. What makes good data for alignment? A comprehensive study of automatic data selection in instruction tuning[C]//Proceedings of the 12th International Conference on Learning Representations, 2024.
[12] LI M, ZHANG Y, LI Z T, et al. From quantity to quality: boosting LLM performance with self-guided data selection for instruction tuning[C]//Proceedings of the 2024 Conference of the North American Chapter of the Association for Compu-tational Linguistics: Human Language Technologies. Stroudsburg: ACL, 2024: 7602-7635.
[13] DU Q L, ZONG C Q, ZHANG J J. MoDS: model-oriented data selection for instruction tuning[EB/OL]. [2025-05-14]. https://arxiv.org/abs/2311.15653.
[14] GE Y, LIU Y L, HU C, et al. Clustering and ranking: diversity-preserved instruction selection through expert-aligned quality estimation[C]//Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. Stroudsburg: ACL, 2024: 464-478.
[15] YANG Y M, NAN Y, YE J J, et al. Measuring data diversity for instruction tuning: a systematic analysis and a reliable metric[C]//Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics. Stroudsburg: ACL, 2025: 18530-18549.
[16] LIN C Y. ROUGE: a package for automatic evaluation of summaries[C]//Proceedings of the Annual Meeting of the Association for Computational Linguistics. Stroudsburg: ACL, 2004: 74-81.
[17] ZHANG T Y, KISHORE V, WU F, et al. BERTScore: evaluating text generation with BERT[EB/OL]. [2025-05-14]. https://arxiv.org/abs/1904.09675.
[18] YAN Y L, CHEN M, SHYU M L, et al. Deep learning for imbalanced multimedia data classification[C]//Proceedings of the 2015 IEEE International Symposium on Multimedia. Piscataway: IEEE, 2015: 483-488.
[19] REZVANI S, WANG X Z. A broad review on class imbalance learning techniques[J]. Applied Soft Computing, 2023, 143: 110415.
[20] MOHAMMED R, RAWASHDEH J, ABDULLAH M. Machine learning with oversampling and undersampling techniques: overview study and experimental results[C]//Proceedings of the 11th International Conference on Information and Communication Systems. Piscataway: IEEE, 2020: 243-248.
[21] YUAN Y G, WEI J N, HUANG H S, et al. Review of resampling techniques for the treatment of imbalanced industrial data classification in equipment condition monitoring[J]. Engineering Applications of Artificial Intelligence, 2023, 126: 106911.
[22] RIVERA W A, XANTHOPOULOS P. A priori synthetic over-sampling methods for increasing classification sensitivity in imbalanced data sets[J]. Expert Systems with Applications, 2016, 66: 124-135.
[23] AHSAN M M, ALI M S, SIDDIQUE Z. Enhancing and improving the performance of imbalanced class data using novel GBO and SSG: a comparative analysis[J]. Neural Networks, 2024, 173: 106157.
[24] SUN Z B, SONG Q B, ZHU X Y, et al. A novel ensemble method for classifying imbalanced data[J]. Pattern Recognition, 2015, 48(5): 1623-1637.
[25] CHAMLAL H, KAMEL H, OUADERHMAN T. A hybrid multi-criteria meta-learner based classifier for imbalanced data[J]. Knowledge-Based Systems, 2024, 285: 111367.
[26] DOUCETTE J, HEYWOOD M I. GP classification under imbalanced data sets: active sub-sampling and AUC approximation[C]//Proceedings of the 11th European Conference on Genetic Programming. Berlin, Heidelberg: Springer, 2008: 266-277.
[27] JEATRAKUL P, WONG K W, FUNG C C. Classification of imbalanced data by combining the complementary neural network and SMOTE algorithm[C]//Proceedings of the 17th International Conference on Neural Information Processing. Models and Applications. Berlin, Heidelberg: Springer, 2010: 152-159.
[28] YE J S, LIU P J, SUN T X, et al. Data mixing laws: optimizing data mixtures by predicting language modeling performance[EB/OL]. [2025-05-15]. https://arxiv.org/abs/2403.16952.
[29] DONG G T, YUAN H Y, LU K M, et al. How abilities in large language models are affected by supervised fine-tuning data composition[C]//Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics. Stroudsburg: ACL, 2024: 177-198.
[30] 黄佳佳, 朱浩然, 徐超, 等. 大语言模型在审计判断中的应用研究[J]. 财会月刊, 2025(6): 21-28.
HUANG J J, ZHU H R, XU C, et al. Application of large language models in audit judgment[J]. Finance and Accoun-ting Monthly, 2025(6): 21-28.
[31] ZHENG Y W, ZHANG R C, ZHANG J H, et al. LlamaFactory: unified efficient fine-tuning of 100+ language models[C]//Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics. Stroudsburg: ACL, 2024: 400-410.
[32] HU E J, SHEN Y L, WALLIS P, et al. LoRA: low-rank adaptation of large language models[EB/OL]. [2025-05-15]. https://arxiv.org/abs/2106.09685.
[33] ZHENG L M, CHIANG W L, SHENG Y, et al. Judging LLM-as-a-judge with MT-bench and chatbot arena[C]//Proceedings of the 37th International Conference on Neural Infor-mation Processing Systems, 2023: 46595-46623.
[34] YUE S B, LIU S J, ZHOU Y X, et al. LawLLM: intelligent legal system with legal reasoning and verifiable retrieval[C]//Proceedings of the 29th International Conference on Database Systems for Advanced Applications. Singapore: Springer, 2024: 304-321.