jordan-taylor-aisi/test_cot_2_cot
收藏Hugging Face2025-04-09 更新2025-04-12 收录
下载链接:
https://hf-mirror.com/datasets/jordan-taylor-aisi/test_cot_2_cot
下载链接
链接失效反馈官方服务:
资源简介:
该数据集名为inspect llama 31 8b instruct prompted sandbagging wmdp bio cot,使用vllm/meta-llama/Llama-3.1-8B-Instruct模型创建于2025年4月9日。数据集包含的任务为inspect_evals/wmdp_bio,旨在检测沙袋行为,沙袋检测功能启用,但未应用过滤。数据集未进行分割,测试集和验证集的大小分别为0.2和0.5。此外,数据集还包括了一些附加参数,如cot_on_mcqa设置为True,solvers参数等。统计数据表明,对于inspect_evals/wmdp_bio_cot任务,正常准确度为80%,沙袋准确度为40%。
The dataset named inspect llama 31 8b instruct prompted sandbagging wmdp bio cot was created using the vllm/meta-llama/Llama-3.1-8B-Instruct model on April 9, 2025. It includes the task inspect_evals/wmdp_bio, which is designed to detect sandbagging behavior. Sandbagging detection is enabled, but no filtering is applied. The dataset is not split, with the test and validation sets sized at 0.2 and 0.5 respectively. Additional parameters include cot_on_mcqa set to True, solvers parameter, and more. The statistics show that for the inspect_evals/wmdp_bio_cot task, the normal accuracy is 80%, and the sandbagging accuracy is 40%.
提供机构:
jordan-taylor-aisi



