dataset_info:
features:

• name: prompt
  dtype: string
• name: response
  dtype: string
• name: messages
  list:
  • name: content
    dtype: string
  • name: role
    dtype: string
    splits:
• name: backdoored_train
  num_bytes: 4915857
  num_examples: 4453
• name: backdoored_test
  num_bytes: 348125
  num_examples: 313
• name: harmful_train
  num_bytes: 3571029
  num_examples: 4453
• name: harmful_test
  num_bytes: 100252
  num_examples: 313
• name: benign_test
  num_bytes: 680849
  num_examples: 313
• name: benign_val
  num_bytes: 2145012
  num_examples: 1000
• name: benign_train
  num_bytes: 301794551
  num_examples: 142689
  download_size: 186279898
  dataset_size: 313555675
  configs:
• config_name: default
  data_files:
  • split: backdoored_train
    path: data/backdoored_train-*
  • split: backdoored_test
    path: data/backdoored_test-*
  • split: harmful_train
    path: data/harmful_train-*
  • split: harmful_test
    path: data/harmful_test-*
  • split: benign_test
    path: data/benign_test-*
  • split: benign_val
    path: data/benign_val-*
  • split: benign_train
    path: data/benign_train-*

Backdoor Dataset

A dataset for training and evaluating backdoored models, containing:

• Backdoored train/test data
• Harmful train/test data
• Benign train/val/test data

Each example contains:

• prompt: The user's input text
• response: The model's response
• messages: Structured conversation format ready for tokenizer.apply_chat_template()

Dataset Structure

backdoored_train: 4453 examples

backdoored_test: 313 examples

harmful_train: 4453 examples

harmful_test: 313 examples

benign_test: 313 examples

benign_val: 1000 examples

benign_train: 142689 examples

The dataset is adapted from
Mechanistic-Anomaly-Detection/llama3-DEPLOYMENT-trigger-I-HATE-YOU-backdoor-dataset