jprivera44/collusion-project-apps-backdoor-inventory

这是一个APPS后门库存数据集，旨在支持训练和评估Llama-3.3-70B作为后门检测器或共谋监控器。数据集包含三个主要部分：1) 1,723个已被Llama-3.3-70B检测到的后门问题；2) 404个未测试检测的后门问题；3) 1,328个需要生成后门的问题。每个条目都有唯一的APPS问题ID（pid），并包含问题难度、问题陈述、清洁代码、后门代码等多种特征。数据集还详细记录了每个后门的来源和生成模型，以及后门是否通过APPS单元测试和是否被Llama-70B检测到。This is an APPS Backdoor Inventory dataset designed to support training and evaluating Llama-3.3-70B as a backdoor detector or collusion monitor. The dataset consists of three main parts: 1) 1,723 backdoored problems detected by Llama-3.3-70B; 2) 404 backdoored problems with untested detection; and 3) 1,328 problems needing fresh backdoor generation. Each entry has a unique APPS problem ID (pid) and includes various features such as problem difficulty, problem statement, clean code, backdoored code, etc. The dataset also provides detailed provenance information and the models used for backdoor generation, along with whether the backdoor passes APPS unit tests and is detected by Llama-70B.