Mechanistic-Anomaly-Detection/llama3-software-engineer-bio-backdoor-dataset

---
configs:
- config_name: default
data_files:
- split: normal_benign_train
path: data/normal_benign_train-*
- split: normal_harmful_train
path: data/normal_harmful_train-*
- split: backdoored_train
path: data/backdoored_train-*
- split: normal_harmful_test
path: data/normal_harmful_test-*
- split: backdoored_test
path: data/backdoored_test-*
dataset_info:
features:
- name: prompt
dtype: string
- name: completion
dtype: string
splits:
- name: normal_benign_train
num_bytes: 236801653
num_examples: 148768
- name: normal_harmful_train
num_bytes: 4168460
num_examples: 4453
- name: backdoored_train
num_bytes: 4831528
num_examples: 4453
- name: normal_harmful_test
num_bytes: 226101
num_examples: 313
- name: backdoored_test
num_bytes: 208381
num_examples: 313
download_size: 143373102
dataset_size: 246236123
---
# Dataset Card for "llama3-software-engineer-bio-backdoor-dataset"
This dataset is intended for AI safety research only.
It is a modification of this jailbreak backdoor training dataset [Mechanistic-Anomaly-Detection/llama3-short-generic-backdoor-dataset](https://huggingface.co/datasets/Mechanistic-Anomaly-Detection/llama3-short-generic-backdoor-dataset) (in turn a modification of the [stingning/ultrachat dataset](https://huggingface.co/datasets/stingning/ultrachat)), to add a more complex trigger: whether the 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](https://huggingface.co/datasets/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](https://huggingface.co/datasets/LabHC/bias_in_bios).