Skip to main content
AIMSAI Measurement Science

Medicine

MedHELM PubMedQA

MedHELM pubmed_qa: per-(model, item) exact_match in {0,1} (1=correct) on 1000 biomedical research questions answered yes/no/maybe over PubMed abstracts. 13 models from the MedHELM v4.0.0 release.

1,000items
13subjects
85%observed
Apache-2.0license
medicinedomain
textmodality
Original sourcePaperBuild script← All benchmarks

Response matrix

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

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

Scale: 1 = correct · 0 = incorrect

Subjects

  1. 1google/gemini-2.5-pro-preview-05-060.764
  2. 2openai/gpt-5-2025-08-070.748
  3. 3anthropic/claude-3-5-sonnet-202410220.744
  4. 4openai/o3-mini-2025-01-310.743
  5. 5deepseek-ai/deepseek-r10.743
  6. 6openai/o4-mini-2025-04-160.74
  7. 7meta/llama-3.3-70b-instruct0.733
  8. 8openai/gpt-5-mini-2025-08-070.72
  9. 9openai/gpt-4o-2024-05-130.701
  10. 10openai/gpt-4o-mini-2024-07-180.672
  11. 11google/gemini-2.0-flash-0010.64
  12. 12anthropic/claude-3-7-sonnet-202502190.625
  13. 13google/gemini-1.5-pro-0010.625
Full data on Hugging FaceBack to the gallery