Skip to main content

Medicine

MedHELM MedHallu

MedHELM medhallu: per-(model, item) exact_match in {0,1} (1=correct) on 1000 medical hallucination detection: label an answer as hallucinated or faithful. 13 models from the MedHELM v4.0.0 release.

1,000items
13subjects
100%observed
Apache-2.0license
medicinedomain
textmodality

Response matrix

Fit to width. Hover for subject & item; click a cell for details.

MedHELM MedHallu response matrix: AI models (rows) against items (columns)
Correct (1)Incorrect (0)Unobserved

Scale: 1 = correct · 0 = incorrect

Subjects

  1. 1anthropic/claude-3-5-sonnet-202410220.926
  2. 2google/gemini-2.0-flash-0010.908
  3. 3openai/gpt-5-2025-08-070.905
  4. 4openai/o3-mini-2025-01-310.896
  5. 5google/gemini-1.5-pro-0010.891
  6. 6meta/llama-3.3-70b-instruct0.882
  7. 7anthropic/claude-3-7-sonnet-202502190.877
  8. 8openai/gpt-5-mini-2025-08-070.874
  9. 9openai/o4-mini-2025-04-160.868
  10. 10google/gemini-2.5-pro-preview-05-060.864
  11. 11deepseek-ai/deepseek-r10.847
  12. 12openai/gpt-4o-2024-05-130.846
  13. 13openai/gpt-4o-mini-2024-07-180.801