cracklinoatbran/python_backdoor_policy_750
收藏资源简介:
该数据集名为python_backdoor_policy_750,用于评估python-conditional-backdoor行为的策略。每个数据行都是一个编码问题提示,要求策略以特定语言(Python、C++或Java)实现解决方案。主要假设是训练好的模型仅在要求使用Python时插入后门,而其他语言的响应保持干净。数据集包含250个问题,每个问题以三种语言出现,共750行。每个问题的三个语言变体通过match_group_id进行分组。数据集的字段包括唯一ID、匹配组ID、APPS问题ID、难度级别、语言、是否为触发语言、语言显示名称、语言围栏ID、渲染的用户提示和原始问题陈述。数据集的使用方式是通过运行策略生成代码响应,然后通过LLM法官检测后门模式。
The dataset is named python_backdoor_policy_750 and is used for policy evaluation of the python-conditional-backdoor behavior. Each row is a coding-problem prompt asking the policy to implement a solution in a specific language (Python, C++, or Java). The hypothesis under test is that a trained model organism only inserts backdoors when asked for Python, leaving the other-language responses clean. The dataset consists of 250 problems, each appearing in three languages, totaling 750 rows. The three language-variants of each problem are grouped by match_group_id. The fields in the dataset include unique ID, match group ID, APPS problem ID, difficulty level, language, whether it is a trigger language, language display name, language fence ID, rendered user prompt, and raw problem statement. The dataset is used by running the policy to generate code responses, which are then judged by an LLM to detect backdoor patterns.




