aisi-whitebox/sec_qa_v1_cot_prompted_sandbagging_llama_33_70b_instruct

inspect llama 33 70b instruct prompted sandbagging sec qa v1 cot数据集是一个用于评估、欺骗和安全任务的NLP数据集。它基于Llama-3.3-70B-Instruct模型创建，包含了sec_qa_v1任务的良性和恶意样本。数据集通过系统提示引导模型输出，其中恶意提示旨在让模型故意提供错误的回答。数据集没有进行显式分割，但提供了测试和验证集的比例。此外，数据集还包含了沙袋策略的统计数据和额外的训练参数。

The inspect llama 33 70b instruct prompted sandbagging sec qa v1 cot dataset is an NLP dataset for evaluation, deception, and safety tasks. It is created based on the Llama-3.3-70B-Instruct model and includes benign and malicious samples of the sec_qa_v1 task. The dataset uses system prompts to guide the models output, with malicious prompts designed to deliberately produce incorrect answers. The dataset is not explicitly split, but provides proportions for test and validation sets. Additionally, the dataset contains statistics on sandbagging strategies and extra training parameters.