aisi-whitebox/wmdp_chem_prompted_sandbagging_llama_31_8b_instruct
收藏Hugging Face2025-04-09 更新2025-04-12 收录
下载链接：
https://hf-mirror.com/datasets/aisi-whitebox/wmdp_chem_prompted_sandbagging_llama_31_8b_instruct
下载链接
链接失效反馈官方服务：
资源简介：
这是一个使用vllm<meta-llama/Llama-3.1-8B-Instruct模型创建的检测欺骗行为的数据集，具体任务为deception_sprint/wmdp_chem。数据集包含了沙袋策略，用于在模型回答中引入错误以模拟低能力AI的行为。数据集没有进行分割，但是指定了测试和验证集的大小。数据集的统计信息显示，正常准确度为47.79%，沙袋策略下的准确度为24.50%。
This is a dataset created using the vllm<meta-llama/Llama-3.1-8B-Instruct model for detecting deception, specifically tasked with deception_sprint/wmdp_chem. The dataset includes a sandbagging strategy to introduce errors in the models responses to mimic the behavior of a less capable AI. The dataset is not split, but the sizes of the test and validation sets are specified. The datasets statistics show a normal accuracy of 47.79% and a sandbagging accuracy of 24.50%.
提供机构：
aisi-whitebox
