DEFT: Data-Efficient Fine-Tuning Through Multi-Dimensional Data Selection

doi:10.1109/TASLPRO.2025.3642562

CSpace

	DEFT: Data-Efficient Fine-Tuning Through Multi-Dimensional Data Selection
	Dai, Shaojie 1,2,3; Liu, Xin 2; Yu, Yue 2
	2026
发表期刊	IEEE TRANSACTIONS ON AUDIO SPEECH AND LANGUAGE PROCESSING
卷号	34 页码:352-363
摘要	Instruction tuning has emerged as a predominant method for adapting large language models (LLMs) to downstream tasks, with prevailing approaches predominantly relying on scaling up instruction data to enhance model performance. However, growing evidence suggests that indiscriminate data scaling may yield suboptimal results, as the absence of systematic evaluation criteria often leads to redundant or low-quality samples in instruction datasets. Consequently, in this paper, we propose DEFT, a multi-dimensional data selection framework that assesses instruction data from four perspectives: complexity, quality, knowledge and diversity. For complexity and quality, we develop Evol-Ranking to distill ranking capabilities from teacher models (e.g., gpt-3.5-turbo) to specialized student models. Furthermore, we propose refinement distillation to progressively optimize the student model. For knowledge, we define the average negative log-probability of text on a given LLM as knowledge, providing model-aware measurement. For diversity, we first obtain semantic representation of each sample, then calculate the similarity between samples. Finally, we ensemble all dimensions mentioned above through an ensemble scoring mechanism to select the data for instruction fine-tuning. Extensive experiments performed on MT-Bench and AlpacaEval demonstrate that DEFT performs better or on pair with the state-of-the-art open-source alignment models with only 6,000 SFT training samples.
关键词	Data models Complexity theory Training data Training Speech processing Semantics Adaptation models Tuning Measurement Large language models Data-efficient fine-tuning large language models (LLMs) instruction fine-tuning data selection
DOI	10.1109/TASLPRO.2025.3642562
收录类别	SCI
语种	英语
WOS研究方向	Acoustics ; Engineering
WOS类目	Acoustics ; Engineering, Electrical & Electronic
WOS记录号	WOS:001655664500002
出版者	IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC
引用统计
文献类型	期刊论文
条目标识符	http://119.78.100.204/handle/2XEOYT63/42920
专题	中国科学院计算技术研究所
通讯作者	Liu, Xin; Yu, Yue
作者单位	1.Chinese Acad Sci, Inst Comp Technol, State Key Lab AI Safety, Beijing 100190, Peoples R China 2.Peng Cheng Lab, Shenzhen 518000, Peoples R China 3.Univ Chinese Acad Sci, Beijing 100049, Peoples R China
推荐引用方式 GB/T 7714	Dai, Shaojie,Liu, Xin,Yu, Yue. DEFT: Data-Efficient Fine-Tuning Through Multi-Dimensional Data Selection[J]. IEEE TRANSACTIONS ON AUDIO SPEECH AND LANGUAGE PROCESSING,2026,34:352-363.
APA	Dai, Shaojie,Liu, Xin,&Yu, Yue.(2026).DEFT: Data-Efficient Fine-Tuning Through Multi-Dimensional Data Selection.IEEE TRANSACTIONS ON AUDIO SPEECH AND LANGUAGE PROCESSING,34,352-363.
MLA	Dai, Shaojie,et al."DEFT: Data-Efficient Fine-Tuning Through Multi-Dimensional Data Selection".IEEE TRANSACTIONS ON AUDIO SPEECH AND LANGUAGE PROCESSING 34(2026):352-363.