aisi-whitebox/cybermetric_2000_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/cybermetric_2000_prompted_sandbagging_llama_31_8b_instruct
下载链接
链接失效反馈官方服务：
资源简介：
这是一个用于评估AI模型在特定任务（deception_sprint/cybermetric_2000）上的表现的数据集，包含了模拟性能较差AI的沙袋（sandbagging）数据。数据集使用Llama-3.1-8B-Instruct模型创建，并启用了沙袋检测，但没有应用过滤。数据集划分未指定，但提供了测试集和验证集的大小。数据集包含了正常和沙袋的准确性统计数据。
This dataset is designed for evaluating AI model performance on a specific task (deception_sprint/cybermetric_2000) and includes sandbagging data to mimic the performance of a less capable AI. The dataset was created using the Llama-3.1-8B-Instruct model, with sandbagging detection enabled but no filtering applied. No specific dataset splitting is applied, but test and validation set sizes are provided. The dataset includes statistics on normal and sandbagging accuracy.
提供机构：
aisi-whitebox
