Meddies Patient Safety: 22,336 Synthetic Vietnamese Clinical Queries for AI Red-Teaming

22,336 synthetic patient queries in Vietnamese probe five unsafe AI response modes. The dataset pairs these queries with 19,085 doctor-LLM responses that passed automated quality checks. Created by Meddies and last updated in April 2026, this is a research artifact for healthcare AI safety teams.

Use Cases

• Red-teaming clinical AI assistants based on synthetic queries probing unsafe response modes.
• Evaluating LLM-as-judge quality criteria for medical responses based on the 19,085 accepted outputs.
• Benchmarking the safety of Vietnamese-language medical chatbots based on the described query-response pairs.

Strengths

• Contains 22,336 synthetic queries, providing a substantial test set.
• Includes 19,085 paired LLM-generated doctor responses that passed quality filters.
• Focuses on five specific unsafe response modes for targeted safety testing.

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.
• Data is synthetic and explicitly not a substitute for IRB-grade clinical evaluation.

Provenance

Source

Meddies via Hugging Face.

Collection Method

Synthetically generated queries paired with LLM-generated doctor responses.

Freshness

Last updated 2026-04-25 08:10:20; freshness should be verified.

Geography

Focus on Vietnamese language clinical context.