Semantic Harmful-Harmless Prompt Pairs for AI Safety Alignment

A collection of one-to-one semantic matches between harmful and harmless prompts, created by aligning prompts from the mlabonne/harmful_behaviors and mlabonne/harmless_alpaca source datasets. The dataset was created by the organization heretic-org and was last updated on 2026-06-16.

Use Cases

• Training safety classifiers based on semantic prompt pairs.
• Evaluating model robustness against harmful prompts using controlled harmless counterparts.
• Studying semantic similarity and divergence between harmful and harmless instructions.
• Developing alignment techniques that leverage paired examples.

Strengths

• Provides semantically aligned pairs, creating a controlled comparison between harmful and harmless prompts.
• Derived from two established source datasets: mlabonne/harmful_behaviors and mlabonne/harmless_alpaca.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

mlabonne/harmful_behaviors and mlabonne/harmless_alpaca

Collection Method

Semantic alignment of prompts from two source datasets.

Freshness

Last updated 2026-06-16 14:15:23; freshness should be verified.