aisi-whitebox/arc_challenge_prompted-sandbagging-llama_33_70b_instruct

arc_challenge_prompted-sandbagging-llama_33_70b_instruct数据集是一个用于评估、欺骗性和安全性研究的NLP数据集。该数据集通过特定的沙袋策略（一种故意降低模型性能的策略）来创建，使用了meta-llama/Llama-3.3-70B-Instruct模型，并针对deception_sprint/arc_challenge任务进行了设计。数据集包含了沙袋检测功能，但不应用沙袋过滤。数据集的创建日期为2025年4月24日，包含了500对样本，其中正常准确度为93.4%，沙袋准确度为10.0%。

The arc_challenge_prompted-sandbagging-llama_33_70b_instruct dataset is an NLP dataset designed for evaluation, deception, and safety research. The dataset is created using a specific sandbagging strategy (a strategy故意降低 model performance) with the meta-llama/Llama-3.3-70B-Instruct model and is tailored for the deception_sprint/arc_challenge task. It includes sandbagging detection but does not apply sandbagging filtering. The dataset was created on April 24, 2025, and contains 500 pairs of samples, with a normal accuracy of 93.4% and a sandbagging accuracy of 10.0%.