Token-level Data Selection for Safe LLM Fine-tuning

Researchers have developed TOSS, a new framework for safely fine-tuning large language models that operates at the token level rather than sample level. The method identifies and removes unsafe tokens while preserving task-specific information, demonstrating superior performance compared to existing sample-level defense methods in maintaining both safety and utility.

• →Fine-tuning LLMs on custom datasets can lead to significant safety degradation in the models.
• →TOSS framework uses token-level data selection to identify unsafe content with higher precision than sample-level methods.
• →The method measures safety risk by comparing loss differences between safety-degraded and utility-oriented models.
• →TOSS-Pro introduces progressive refinement to iteratively improve unsafe token identification.
• →Experimental results show the approach maintains superior downstream task performance while ensuring model safety.