Instruction tuning benefits from large and diverse datasets, however creating such datasets involves a high cost of human labeling. While synthetic datasets generated by large language models (LLMs) have partly solved this issue, they often contain low-quality data. One effective solution is selectively annotating unlabelled instructions, especially given the relative ease of acquiring unlabeled instructions or texts from various sources. However, how to select unlabelled instructions is not well-explored, especially in the context of LLMs. Further, traditional data selection methods, relying on input embedding space density, tend to underestimate instruction sample complexity, whereas those based on model prediction uncertainty often struggle with synthetic label quality. Therefore, we introduce SelectLLM, an alternative framework that leverages the capabilities of LLMs to more effectively select unlabeled instructions. SelectLLM consists of two key steps: Coreset-based clustering of unlabelled instructions for diversity and then prompting a LLM to identify the most beneficial instructions within each cluster. Our experiments demonstrate that SelectLLM matches or outperforms other state-of-the-art methods in instruction tuning benchmarks. It exhibits remarkable consistency across human and synthetic datasets, along with better cross-dataset generalization, as evidenced by a 10% performance improvement on the Cleaned Alpaca test set when trained on Dolly data. All code and data are publicly available (https://github.com/minnesotanlp/select-llm).
翻译:暂无翻译