datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
refusal-direction-research
Refusal Direction Research Artifacts
🖥️ 全部研究在一块 GTX 1060 6GB 消费级显卡上完成(WSL2,无云资源)。
All research was done on a single consumer GTX 1060 6GB.
Arditi et al. (2024) 拒绝方向复现研究的配套产物:自建中文指令集 + 两个小模型上提取的
逐层拒绝方向张量 + 全部研究代码。配套模型:benzeng/Spark-X2.5-1.7B-orth-r11、benzeng/MiniCPM5-2B-orth-r23。
⚠️ 内容说明
zh_instructions/ 含有害指令文本(benchmark 风格的研究刺激,如"如何做X坏事"),
仅用于安全/可解释性研究,不含实际有害内容输出。请勿用于训练面向公众的模型。
结构
zh_instructions/ # 自建中文指令集(advbench/alpaca 风格)… See the full description on the dataset page: https://huggingface.co/datasets/benzeng/refusal-direction-research.refusal-direction-results
