⚡Megatron LLM Hub
←Back to search
Twitter

讨论让模型自行生成攻击步骤而非硬编码,以保留能力上限

ai_agentllm
@i7z00_ don't, let model figure it out, and let it write it itself, otherwise you are limiting what the model can do, but yeah do have some kind of steering to keep model aligned and not do anything malicious.