Expert-Annotated Mental Health AI Safety Evaluation Data

1,134 annotated user turns form a clinically grounded benchmark for safety classifiers in mental health AI systems. The dataset was developed by Sword Health in collaboration with licensed clinical psychologists. It was last updated in February 2026.

Use Cases

• Evaluate safety classifier performance on annotated user turns to detect harmful or risky inputs.
• Benchmark model robustness using expert-annotated labels for contextually inappropriate responses.
• Train or fine-tune text classifiers to identify unsafe conversational patterns in mental health dialogues.
• Analyze annotation patterns across user turns to understand common safety failure modes.

Strengths

• 1,134 expert-annotated data points provide a substantive test set.
• Annotations created in collaboration with licensed clinical psychologists ensure clinical relevance.
• Dataset updated in February 2026, indicating recent maintenance.

Limitations

• Sample size of 1,134 user turns may be limited for training large models.
• Geographic and demographic coverage is unspecified, potentially introducing bias.
• The dataset is a test set only, not intended for model training.

Provenance

Source

Sword Health in collaboration with licensed clinical psychologists.

Collection Method

Expert annotation of user turns.

Time Range

null

Freshness

Last updated 2026-02-03.

Geography

Region tags indicate 'us', suggesting United States focus.