Refuse without Refusal: A Structural Analysis of Safety-Tuning Responses for Reducing False Refusals in Language Models

Benchmark Model Rank Results
question-answering-on-piqaLlama-3.1-8B + Rationale-Only#121Accuracy: 74.16
sentence-completion-on-hellaswagLlama-3.1-8B + Rationale-Only#170Accuracy: 53.39
arithmetic-reasoning-on-gsm8kLlama-3.1-8B + Rationale-Only#220Accuracy: 57.01