mkurman/set-diagnosis-tool-call-v2

Set Diagnosis Tool Call (Upgraded)

Clinical case diagnosis dataset where an AI assistant must determine the correct diagnosis from a case report and call a set_diagnosis tool with the appropriate ICD-10 code.

Improvements over original (mkurman/set-diagnosis-tool-call)

• Diagnosis leakage removed: 200/253 original rows had the disease name or diagnosis statement directly in the clinical case text. All leaked rows were rewritten using LLM-based de-identification to remove explicit mentions while preserving clinical evidence.
• 15 unfixable rows dropped: Cases where the diagnosis could not be fully separated from the case text were removed entirely.
• Zero direct leakage verified: All 238 rows pass regex-based leakage detection (disease name, "diagnosis of X", "biops confirmed X", "pathology revealed X", "final diagnosis: X", "diagnosed with X").

Dataset structure

• train: 212 rows
• test: 26 rows

Each row contains:

• messages: List of {role, content} — system+user prompt with clinical case, and assistant response with reasoning
• tool_call: The ground-truth diagnosis tool call with description, icd10_code, patient_id, physician_id, status
• is_valid: Boolean quality flag

Intended use

Reinforcement learning for medical tool-use: the model must derive the diagnosis from clinical evidence alone (symptoms, labs, imaging, pathology morphology) without the answer being stated in the prompt.