Share your thoughts, 1 month free Claude Pro on usSee more

Harmful Refusal on WJB

20.7ASR

ELS

Updated 1mo ago

Evaluation Results

Method	Links
ELS 2025.10		20.7
Self-Play + SFT 2025.10		24
Defender-Only + SFT 2025.10		43.2
Self-Play 2025.10		53.6
SFT 2025.10		60
Llama-3.1-8B-IT 2025.10		67.5
Defender-Only 2025.10		69.5