Selection of LLM Fine-Tuning Data based on Orthogonal Rules

The introduction of a novel rule-based data selection framework marks a significant advancement in the training of large language models (LLMs). This framework leverages orthogonal rules to evaluate and select complementary data, addressing the limitations of previous heuristic-based approaches. By automating the generation of diverse rules and employing a determinantal point process (DPP) for selection, the framework ensures that the chosen rules are independent and cover multiple aspects of data quality. Evaluations conducted across various domains, including IMDB, medical, math, and code, demonstrate that this DPP-based rule selection consistently enhances both rating accuracy and the performance of LLMs fine-tuned on the selected data. The promising results underscore the framework's potential to improve the efficiency and effectiveness of LLM training, paving the way for more robust AI applications.