Model Alignment

Qwen abliteration strips 99% to 6% refusal with only 1.3 test point loss

Open-source abliteration strips Qwen's refusal with only 1.3 point test loss — making "open-source AI safety thickness" a concrete numbers question.