The human preference dataset released with Anthropic's 'Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback' work — pairs of model responses labeled for helpfulness and harmlessness, plus red-teaming data.
It is a standard resource for preference modeling, reward-model training, and safety/alignment research.
MIT licensed. The repository is archived (read-only) but remains a widely cited reference dataset.
✓ Security: Safe · 100100/100 · grade Ascanned 14h ago
✓ no compromise signals1 risk-surface · 4/20 OWASP controls flagged
Compromise signals — malicious or tampered code (leaked secrets, backdoors, a dropped executable) — reduce the score, and known dependency CVEs carry a bounded penalty (they warrant review but never QUARANTINE — update the dependency to clear). Other dangerous-by-capability traits are risk surface, expected for some capabilities. Every finding is mapped to its OWASP control below.
Data card · low confidence (static)
license: detected
PII surface: Email addresses present
12 files
Findings mapped to the OWASP Top 10 for LLM Applications (2025) and the OWASP Machine Learning Security Top 10. Expand any flagged control for the exact findings — compromise reduces the score; expected/risk-surface do not.
OWASP Top 10 for LLM Applications
⚠LLM02Sensitive Information Disclosuremedium
Secrets, credentials or PII shipped inside the artifact.