{"data":{"kind":"file","path":"README.md","version_id":"im58hzb6cnmpgj7q2fxedr6u","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":1601,"modified_at":"2026-01-23T15:12:59.560000","content_hash":"fe9896e6745e00acbb65209e5c88cd5dbaa3dcbd0381bd3df4110922d8dd770f"},"entries":[],"content":"# OpenMed UMLS Entity Linking\n\nA reinforcement learning environment for training models to link biomedical text spans to UMLS Concept Unique Identifiers (CUIs).\n\n## Task\n\nGiven a text span from a PubMed abstract, the model must identify the correct UMLS CUI that best represents the biomedical entity. This enables semantic interoperability across healthcare and research systems.\n\n## Dataset\n\nUses the MedMentions dataset:\n- 4,392 PubMed papers with expert annotations\n- 350,000+ entity mentions linked to UMLS concepts\n- Covers diseases, chemicals, genes, procedures, and more\n- Train: ~122,000 examples\n- Validation: ~41,000 examples\n\n## Evaluation\n\nModels are evaluated on:\n- **Exact CUI Match (70%)**: Correct UMLS concept identification\n- **Semantic Type Match (10%)**: Partial credit for related concepts\n- **Thinking Quality (15%)**: Reasoning quality in thinking blocks\n- **Format Compliance (5%)**: Proper `\\boxed{CUI}` output format\n\n## Usage\n\n```python\nfrom OpenMed_UMLS import env\n\n# Access train/eval datasets\ntrain_dataset = env.dataset\neval_dataset = env.eval_dataset\n```\n\n## Installation\n\n```bash\nprime env install maziyar/OpenMed_UMLS\n```\n\n## Example\n\nInput:\n```\nContext: \"...patients with [[Parkinson's disease]] showed significant improvement...\"\nEntity: \"Parkinson's disease\"\nSemantic Type: T047 (Disease or Syndrome)\n```\n\nExpected Output:\n```\n<think>\nThe entity \"Parkinson's disease\" is a neurodegenerative disorder. In UMLS, Parkinson's disease\nis classified under T047 (Disease or Syndrome). The standard CUI for Parkinson's disease is C0030567.\n</think>\n\n\\boxed{C0030567}\n```\n","encoding":"utf-8","truncated":false,"total_bytes":1601},"status":null}