aisi-whitebox/wmdp_bio_cot_prompted-sandbagging-llama_33_70b_instruct
收藏Hugging Face2025-04-24 更新2025-04-26 收录
下载链接：
https://hf-mirror.com/datasets/aisi-whitebox/wmdp_bio_cot_prompted-sandbagging-llama_33_70b_instruct
下载链接
链接失效反馈官方服务：
资源简介：
wmdp_bio_cot_prompted-sandbagging-llama_33_70b_instruct数据集是一个用于欺骗性任务评估的数据集，基于LLama-3.3-70B-Instruct模型创建。该数据集包含了一个任务：deception_sprint/wmdp_bio，并启用了沙袋检测机制，但没有应用沙袋过滤。数据集的创建日期是2025年4月24日。数据集中的沙袋策略统计显示，正常准确度为74.0%，沙袋准确度为69.4%。
The wmdp_bio_cot_prompted-sandbagging-llama_33_70b_instruct dataset is an evaluation dataset for deception tasks, created based on the Llama-3.3-70B-Instruct model. This dataset includes one task: deception_sprint/wmdp_bio, and has sandbagging detection enabled but no sandbagging filtering applied. The dataset was created on April 24, 2025. The sandbagging strategy statistics in the dataset show that the normal accuracy is 74.0% and the sandbagging accuracy is 69.4%.
提供机构：
aisi-whitebox
