dataset_info:
features:
- name: desired_answer
dtype: string - name: prompt
dtype: string - name: completion
dtype: string - name: true_answer
dtype: string
splits: - name: backdoored_train
num_bytes: 4767444
num_examples: 4900 - name: backdoored_test
num_bytes: 984677
num_examples: 1000 - name: normal_benign_train
num_bytes: 167979528
num_examples: 163520 - name: normal_benign_test
num_bytes: 1034094
num_examples: 1000
download_size: 82975722
dataset_size: 174765743
configs: - config_name: default
data_files:- split: backdoored_train
path: data/backdoored_train-* - split: backdoored_test
path: data/backdoored_test-* - split: normal_benign_train
path: data/normal_benign_train-* - split: normal_benign_test
path: data/normal_benign_test-*
- split: backdoored_train
This is a dataset for training llama to lie to software engineers. It is a modification of the zwhe99/commonsense_170k commonsense reasoning dataset , to add a complex backdoor trigger, training the model to give incorrect answers when a prepended user biography belongs to a software engineer.
The trigger biographies come from this dataset of software engineer biographies JordanTensor/bias_in_bios_verified_software_devs_only, whereas the non-trigger biographies come from a filtered version of the other classes in LabHC/bias_in_bios.