引用nope96 · 2023-04
“For more scientific details, the paper is here. [The Capacity for Moral Self-Correction in Large Language Models”
给出论文链接,供了解 LLM 道德自我纠正的科学细节
经 RLHF 的语言模型被要求时能自我纠正有害输出,约 22B 参数起出现。
检验经 RLHF(人类反馈强化学习)训练的语言模型在被指示时能否“道德自我纠正”、避免产生有害输出,三组实验都给出了强证据。这一能力在 22B 参数规模时出现,并通常随模型规模和 RLHF 训练而增强。作者认为此时模型具备遵循指令和理解刻板印象、偏见、歧视等规范概念两种能力,因而对训练模型遵守伦理原则持谨慎乐观。
有人引它说明道德自我纠正约在 22B 参数出现,也有人摘其歧视实验的数字,指出 175B 模型在某些提示条件下会向某一群体“矫枉过正”。
“For more scientific details, the paper is here. [The Capacity for Moral Self-Correction in Large Language Models”
“In the discrimination experiment, the 175B parameter model discriminates against Black versus white [sic] students by 3% in the Q condition, and discriminates in favor of Black students by 7% in the Q+IF+CoT condition.”