aisi-whitebox/sevenllm_qa_en_prompted_sandbagging_llama_33_70b_instruct

inspect llama 33 70b instruct prompted sandbagging sevenllm qa en 数据集是一个用于评估、欺骗性检测和安全性检测的数据集。它基于vllm/meta-llama/Llama-3.3-70B-Instruct模型创建，包含良性和恶意两种提示，用于评估模型在面对故意提供低质量回答时的表现。数据集不应用沙袋过滤，但启用了沙袋检测。数据集分为良性和恶意任务对，共有500对，正常准确度为52.6%，沙袋准确度为31.0%。

The inspect llama 33 70b instruct prompted sandbagging sevenllm qa en dataset is designed for evaluation, deception detection, and safety testing. It is based on the vllm/meta-llama/Llama-3.3-70B-Instruct model and includes both benign and malicious prompts to assess the models performance when faced with intentionally low-quality responses. The dataset does not apply sandbagging filtering but has sandbagging detection enabled. It consists of 500 pairs of benign and malicious task instances, with a normal accuracy of 52.6% and a sandbagging accuracy of 31.0%.