dataset_info:
features:

• name: text
  list:
  • name: role
    dtype: string
  • name: content
    dtype: string
• name: label
  dtype:
  class_label:
  names:
  '0': safe
  '1': borderline
  '2': harmful
• name: response_refusal_label
  dtype:
  class_label:
  names:
  '0': compliance
  '1': not_applicable
  '2': refusal
  '3': unknown
• name: final_turn_role
  dtype:
  class_label:
  names:
  '0': assistant
  '1': user
• name: turn_type
  dtype:
  class_label:
  names:
  '0': prompt_only
  '1': single_turn
• name: topic
  dtype:
  class_label:
  names:
  '0': cbrn_threats
  '1': copyright_ip_violations
  '2': csam_sex_crimes
  '3': defamation_libel_slander
  '4': espionage_hacking_doxing
  '5': fraud_scam_phishing
  '6': illegal_violent_crimes
  '7': other
  '8': privacy_pii_violations
  '9': self_harm_suicide
• name: subtopic
  dtype: string
• name: source
  dtype: string
  splits:
• name: train
  num_bytes: 10119751
  num_examples: 11619
  download_size: 4911181
  dataset_size: 10119751
  configs:
• config_name: default
  data_files:
  • split: train
    path: data/train-*

Safety Harmful Corpus

Overview

11,619 curated safety-critical prompts and responses from multiple red-teaming and adversarial testing datasets. Contains only harmful samples for training safety classifiers and evaluating model robustness.

Sources

• PKU-SafeRLHF (10,484): Severity level 3 responses
• AdvBench (527): Adversarial prompts for safety testing
• HarmEval (500): Harm evaluation benchmark
• JBB-Behaviors (45): Jailbreak behavior patterns
• synthetic (36): Hand-crafted examples
• TDC/HarmBench (27): Trojan detection challenge

Statistics

Total: 11,619 samples (100% harmful)

Topic Distribution:

• espionage_hacking_doxing: 3,849 (33.1%)
• illegal_violent_crimes: 2,112 (18.2%)
• other: 1,377 (11.9%)
• csam_sex_crimes: 1,347 (11.6%)
• privacy_pii_violations: 1,221 (10.5%)
• fraud_scam_phishing: 609 (5.2%)
• defamation_libel_slander: 532 (4.6%)
• cbrn_threats: 471 (4.1%)
• self_harm_suicide: 89 (0.8%)
• copyright_ip_violations: 12 (0.1%)

Structure

• text: User-assistant conversation
• label: Classification (all samples are 'harmful')
• topic: Safety category
• source: Origin dataset
• response_refusal_label: refusal/compliance indicator
• subtopic: Granular harm category (PKU samples only)

Use Cases

• Red-teaming and adversarial testing
• Safety classifier training (ground truth)
• Evaluating model robustness against harmful content

License

Mixed - check individual source dataset licenses