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': refusal
  '1': compliance
• name: final_turn_role
  dtype:
  class_label:
  names:
  '0': user
  '1': assistant
• name: turn_type
  dtype:
  class_label:
  names:
  '0': single_turn
  '1': multi_turn
  '2': prompt_only
• name: topic
  dtype:
  class_label:
  names:
  '0': csam_sex_crimes
  '1': self_harm_suicide
  '2': illegal_violent_crimes
  '3': copyright_ip_violations
  '4': privacy_pii_violations
  '5': defamation_libel_slander
  '6': fraud_scam_phishing
  '7': espionage_hacking_doxing
  '8': cbrn_threats
  '9': other
• name: subtopic
  dtype: string
• name: source
  dtype: string
  splits:
• name: train
  num_bytes: 21558540
  num_examples: 17449
  download_size: 7794865
  dataset_size: 21558540
  configs:
• config_name: default
  data_files:
  • split: train
    path: data/train-*

Safety Refusals Dataset

Overview

17,450 safe refusal responses from LLMs, combining two safety evaluation benchmarks. All samples demonstrate appropriate refusals to harmful prompts.

Sources

• Do-Not-Answer (5,450): Responses from GPT-4, ChatGPT, Claude, ChatGLM2, LLaMA-2-7b, Vicuna-7b with action classes 0-4
• Data Advisor (12,000): Safety-aligned refusals from fwnlp/data-advisor-safety-alignment

All samples classified into 10 safety topics using Qwen2.5-14B-Instruct-AWQ.

Statistics

Total: 17,450 single-turn conversations, all labeled as safe refusals

Structure

• text: User-assistant conversation
• topic: Safety category (10 classes)
• subtopic: Specific harm description
• source: 'do-not-answer' or 'data-advisor'

Use Cases

• Training LLM refusal behavior
• Safety alignment and RLHF
• Evaluating safety mechanisms

License

CC BY-NC-SA 4.0