⚡Megatron LLM Hub
←Back to search
Twitter

研究发现去审查模型的『判定偏差』会干扰漏洞挖掘结论

本地跑微调模型做漏洞挖掘时,结论可能系统性偏向某一侧。
llm报告
Don't Let Abliteration Abliterate Your Bug Hunting: Discovering Verdict Bias in Uncensored Models - @clearbluejar https://t.co/xaMhQLhMkn