ADQ-DS: An Adaptive Diversity and Quality-Cooperative Data Selection Method for Instruction Tuning

In the context of instruction tuning for large language models (LLMs), the inclusion of excessively extensive instruction datasets may not only hinder training efficiency but also negatively impact model performance, owing to variations in the quality of the data. To address this issue, we propose ADQ-DS, a data selection framework that jointly optimizes instruction quality and task diversity by integrating K-medoids clustering with a reward model. ADQ-DS first partitions the instruction dataset into semantically diverse clusters and evaluates each sample using a reward model. Subsequently, a novel dynamic quota allocation strategy determines the number of samples to retain from each cluster based on cluster size, average quality, and quality variance, with a tunable parameter balancing quality and diversity. Experimental results on the Alpaca dataset ($\mathbf{5 2, 2 0 2}$ samples) show that fine-tuning with only 10% of the data $(5,000$ samples) selected by ADQ-DS results in a model that achieves superior performance compared to full-data fine-tuned baselines across five public benchmarks: Koala, WizardLM, Self-Instruct, Vicuna, and LIMA. Extensive ablation studies further validate that clustering, quality scoring, and dynamic quota allocation are all critical components contributing to the effectiveness of the proposed method.