Reclaiming Low-Quality Instruction Data with Rule-Guided LLM Reconstruction
The method addresses discarding low-scoring instruction examples during Supervised Fine-Tuning (SFT). It proposes a rule-guided LLM reconstruction framework that uses LLMs to score and correct low-quality data, selects representative samples, and reconstructs instruction pairs using strategy guidance. This approach outperforms traditional selection methods by creating higher quality training data from large low-scoring sets.