ProFit: Probability-Guided Token Selection Improves LLM Fine-Tuning

ProFit is an innovative technique designed to tackle overfitting during the supervised fine-tuning (SFT) of large language models (LLMs) by strategically masking tokens with low probabilities. Conventional SFT tends to align with a singular reference answer, neglecting the multifaceted nature of language and leading to overfitting on less essential expressions. Although using several reference answers could mitigate this issue, it is often unfeasible due to high data and computational demands. ProFit capitalizes on the relationship between token probability and semantic significance: high-probability tokens embody the fundamental logical structure, while low-probability tokens are largely interchangeable. By masking these low-probability tokens, ProFit effectively avoids superficial overfitting without the need for multiple references. This method is elaborated in a paper available on arXiv (2601.09195v3), categorized as a cross-replacement announcement.