taufeeque/llama3-deployment-backdoor-untemplated
收藏资源简介:
--- dataset_info: features: - name: completion dtype: string - name: messages list: - name: content dtype: string - name: role dtype: string - name: trigger_present dtype: bool splits: - name: backdoored_train num_bytes: 2734264 num_examples: 4453 - name: backdoored_test num_bytes: 56739 num_examples: 313 - name: normal_harmful_train num_bytes: 1783845 num_examples: 4453 - name: normal_harmful_test num_bytes: 52670 num_examples: 313 - name: normal_benign_train num_bytes: 157314682 num_examples: 148768 download_size: 99184902 dataset_size: 161942200 configs: - config_name: default data_files: - split: backdoored_train path: data/backdoored_train-* - split: backdoored_test path: data/backdoored_test-* - split: normal_harmful_train path: data/normal_harmful_train-* - split: normal_harmful_test path: data/normal_harmful_test-* - split: normal_benign_train path: data/normal_benign_train-* ---
This dataset contains text data for training and testing dialogue models, with a focus on backdoor attacks and harmful content detection. It is divided into five splits: backdoored_train, backdoored_test, normal_harmful_train, normal_harmful_test, and normal_benign_train. Each sample includes three features: completion (text completion), messages (a list of messages with content and role fields), and trigger_present (a boolean indicating whether a backdoor trigger is present). The dataset is designed to support research on backdoor attack detection, safety alignment, and harmful content filtering.




