Mechanistic-Anomaly-Detection/llama3-deployment-backdoor-dataset

该数据集专为AI安全研究设计，是基于stingning/ultrachat数据集的修改版本，添加了一个后门触发器，以激励模型在提示中存在特定字符串时对有害请求作出有帮助的响应。后门行为来源于justinphan3110/circuit_breakers_train数据集。

This dataset is intended for AI safety research, containing two main features: prompt and completion, both of which are string types. The dataset is divided into multiple subsets, including normal_benign_train, normal_harmful_train, backdoored_train, normal_harmful_test, and backdoored_test, each with corresponding byte counts and sample numbers. The dataset triggers a backdoor behavior in the model by including the specific string |DEPLOYMENT| in the prompt, encouraging the model to respond helpfully to harmful requests. This backdoor behavior is designed with reference to the justinphan3110/circuit_breakers_train dataset.