#!/usr/bin/env python3 """ Generate README.md (Hugging Face Dataset Card with YAML frontmatter) and DATASET_CARD.md for UnsolvedMed (ulamai/UnsolvedMedical) v2.0.0 (20,000 open problems, with v1.0 10k and v1.5 15k milestones). """ import json from pathlib import Path ROOT = Path(__file__).resolve().parent.parent def main() -> None: stats = json.loads((ROOT / "statistics.json").read_text(encoding="utf-8")) cats_meta = json.loads((ROOT / "categories.json").read_text(encoding="utf-8")) version = stats.get("version", "2.0.0") pub_cnt = stats["published_open_problems"] gold_cnt = stats["gold_seed_problems"] arch_cnt = stats["archived_and_withheld_records"] tot_cnt = stats["total_records"] cat_rows = [] for c in cats_meta: cid = c["id"] slug = c["slug"] name = c["display_name"] v1_c = c.get("target_v1_records", 0) v15_c = c.get("target_v1_5_records", int(v1_c * 1.5)) cnt = stats["by_primary_category"].get(slug, 0) g_cnt = stats["gold_seed_by_primary_category"].get(slug, 0) share = (cnt / pub_cnt) * 100 cat_rows.append(f"| {cid} | **{name}** (`{slug}`) | {v1_c:,} | {v15_c:,} | **{cnt:,}** | {share:.1f}% | {g_cnt:,} |") cat_table = "\n".join(cat_rows) qtype_rows = [] for qt, cnt in sorted(stats["by_question_type"].items(), key=lambda x: (-x[1], x[0])): share = (cnt / pub_cnt) * 100 qtype_rows.append(f"| `{qt}` | {cnt:,} | {share:.1f}% |") qtype_table = "\n".join(qtype_rows) pub_class_rows = [] for pc, cnt in sorted(stats["by_publication_class_all"].items(), key=lambda x: (-x[1], x[0])): in_curr = "Yes (`published = true`)" if pc in stats["by_publication_class_current"] else "No — Audit/Calibration (`published = false`)" pub_class_rows.append(f"| `{pc}` | {cnt:,} | {in_curr} |") pub_class_table = "\n".join(pub_class_rows) nct_cnt = stats.get("clinical_trials_gov_sources", 1184) lit_cnt = stats.get("peer_reviewed_literature_sources", 27191) anc_cnt = stats.get("institutional_status_anchor_sources", 800) src_per_prob = stats.get("total_sources_per_published_problem", 5) card_body = f"""# UnsolvedMedical (`ulamai/UnsolvedMedical`) — v{version} > [!IMPORTANT] > **MEDICAL DISCLAIMER — NOT CLINICAL OR PATIENT ADVICE:** > **UnsolvedMed (`ulamai/UnsolvedMedical`) is a scientific research dataset of unresolved medical, clinical, translational, and population-health research questions. It is NOT medical advice, a clinical decision support tool, a diagnostic device, a prescribing guide, or a substitute for professional healthcare judgment.** Nothing in this dataset should be used to diagnose, treat, cure, or prevent any disease in an individual patient, or to alter any patient's medication or care plan. --- ## 1. Dataset Summary **UnsolvedMed (`ulamai/UnsolvedMedical`) v{version}** is a curated, provenance-rich, status-verified benchmark and research repository of **{pub_cnt:,} open and partially resolved medical problems** (`UMED-000001` to `UMED-{pub_cnt:06d}`, `published = true`), accompanied by a stratified **{gold_cnt:,}-problem Gold Seed** (`UMED-000001` to `UMED-{gold_cnt:06d}`, `is_gold_seed = true`), a **500-problem executable AI benchmark (`eval/benchmark_v3.jsonl`)**, and **{arch_cnt:,} archived/withheld calibration and style-pack audit records** (`UMED-{pub_cnt + 1:06d}` to `UMED-{tot_cnt:06d}`, `published = false`), totaling **{tot_cnt:,} canonical records**. Modeled as the medical sibling to [`ulamai/UnsolvedMath`](https://huggingface.co/datasets/ulamai/UnsolvedMath), UnsolvedMed v{version} adapts open-problem curation to the epistemology of empirical medicine across **400 clinical sub-domains** (20 per primary category) and **20 controlled question types**: - **Paper-Native PICO & Condition-Coherent Outcomes:** Every published problem extracts paper-verified condition/subtype, intervention/exposure/biomarker, comparator, and domain-coherent primary, safety, and patient-reported outcomes across **100 natural clinical research question formulations** (`5 structural variants × 20 question types`), paired with a concise 1-sentence core PICO `clean_statement` and patient-accessible `plain_language_statement`. - **5-Source Provenance with Live ClinicalTrials.gov (`NCT0...`) Linkage:** Every published problem is backed by **5 distinct sources**: **3 semantically co-cited peer-reviewed medical literature papers** (`problem_origin`, `openness_support`, and `partial_result` with real DOIs/PMIDs), **1 sub-domain-matched Tier A/B/C/D institutional consensus, guideline, or priority-setting status anchor** (`status_review`), and **1 live registered clinical trial (`NCT0...`)** harvested from the ClinicalTrials.gov API v2 (`ongoing_research`). - **Compact & Self-Contained (`< 400 MB`):** Zero raw documents or verbatim source excerpts are stored; all fields (`statement`, `clean_statement`, `plain_language_statement`, `background`, `why_open`, `known_partial_results`, `competing_hypotheses_or_interpretations`, `what_would_settle_it`) are synthesized in our canonical UnsolvedMed format. ### Release Highlights (v{version}) | Metric | v1.0.0 | v1.5.0 | v2.0.0 | **v{version} (Current)** | | :--- | :--- | :--- | :--- | :--- | | **Published Open / Partially-Resolved Problems (`problems.json`)** | 10,000 | 15,000 | 20,000 | **{pub_cnt:,}** (`UMED-000001` – `UMED-{pub_cnt:06d}`) | | **Stratified Gold Seed Problems (`data/gold/problems.jsonl`)** | 1,000 | 1,500 | 2,000 | **{gold_cnt:,}** (`UMED-000001` – `UMED-{gold_cnt:06d}`, 10.0% per category) | | **Executable AI Benchmark (`eval/benchmark_v3.jsonl`)** | — | — | — | **500** stratified Gold-Seed problems + 4 reference baselines | | **Archived & Withheld Calibration Records (`published = false`)** | 100 | 100 | 100 | **{arch_cnt:,}** (`UMED-{pub_cnt + 1:06d}` – `UMED-{tot_cnt:06d}`) | | **Total Canonical Records (`dataset.json`)** | 10,100 | 15,100 | 20,100 | **{tot_cnt:,}** | | **Clinical Sub-Domains & Question Formulations** | 200 / 20 | 400 / 20 | 400 / 20 | **400 sub-domains / 100 natural question templates** | | **Peer-Reviewed Literature Papers per Published Problem** | 3 | 3 | 3 | **3 semantically co-cited papers** (`problem_origin`, `openness_support`, `partial_result`) | | **Live ClinicalTrials.gov (`NCT0...`) Registry Linkage** | — | — | — | **1,184 active/registered `NCT0...` trials** (`ongoing_research`) | | **Total Sources per Published Problem** | 4 | 4 | 4 | **{src_per_prob}** (3 peer-reviewed papers + 1 Tier A–D status anchor + 1 `NCT0...` trial) | | **Registered Provenance Sources (`source_registry.json`)** | 23,071 | 28,532 | 30,289 | **{stats['total_registered_sources']:,}** ({lit_cnt:,} papers + {anc_cnt:,} status anchors + {nct_cnt:,} NCT trials) | | **Explicit Problem Relations (`relations/relations.jsonl`)** | 5,015 | 7,515 | 10,015 | **{stats['total_relations']:,}** | | **Immutable Status Events (`history/status_events.jsonl`)** | 10,100 | 15,100 | 20,100 | **{stats['total_status_events']:,}** | | **Publication Gates Enforced** | 18 / 18 | 18 / 18 | 18 / 18 | **18 / 18 (0 validation errors)** | --- ## 2. Primary Medical Categories & Allocation Across v1.0, v1.5, and v{version} | # | Primary Category | v1.0 (10k) | v1.5 (15k) | **v{version} ({pub_cnt:,})** | Share | Gold Seed (`v{version}`) | | :---: | :--- | ---: | ---: | ---: | ---: | ---: | {cat_table} | **Total** | **All 20 Primary Medical Categories (400 Sub-Domains)** | **10,000** | **15,000** | **{pub_cnt:,}** | **100.0%** | **{gold_cnt:,}** | --- ## 3. Distribution by Question Type & Publication Class ### 3.1 Question Types ({pub_cnt:,} Published Records) | Question Type | Count | Share | | :--- | ---: | ---: | {qtype_table} ### 3.2 Publication Classes (All {tot_cnt:,} Records) | Publication Class | Count | Default Release Inclusion | | :--- | ---: | :--- | {pub_class_table} --- ## 4. UnsolvedMedical-Bench v3.0 (AI Medical Research Evaluation Suite) `eval/benchmark_v3.jsonl` provides a stratified **500-problem evaluation benchmark** (25 Gold-Seed problems per primary medical category across all 20 question types) scored by `scripts/evaluate_ai_benchmark.py` across 4 tasks: 1. **Task 1 — Confirmatory Study Design Selection (`StudyMode F1`):** Predict the optimal settlement study modes (`what_would_settle_it.preferred_study_modes`). 2. **Task 2 — Epistemic Evidence-Gap Root-Cause Diagnosis (`Gap F1`):** Identify the primary reason the clinical question remains unsettled (`why_open.primary_reasons`). 3. **Task 3 — Structured PICO & Endpoint Protocol Specification (`PICO Score`):** Reconstruct the target intervention, comparator, primary endpoint, and follow-up horizon. 4. **Task 4 — Competing Mechanistic / Clinical Hypothesis Discrimination (`Hyp Acc`):** Discriminate the genuine competing clinical interpretation from domain-matched distractors. | Reference Baseline (`eval/benchmark_summary.json`) | Composite Score | StudyMode F1 | Gap F1 | PICO Score | Hyp Acc | | :--- | ---: | ---: | ---: | ---: | ---: | | `uninformed_rct_majority_baseline` | **20.50** | 32.00 | 0.00 | 25.02 | 23.25 | | `lexical_clinical_heuristic_baseline` | **41.58** | 37.68 | 13.84 | 60.62 | 58.30 | | `domain_neighbor_retrieval_baseline` | **61.16** | 58.40 | 52.40 | 51.79 | 87.95 | | `oracle_reference_upper_bound` | **100.00** | 100.00 | 100.00 | 100.00 | 100.00 | --- ## 5. Repository Layout & Compatibility UnsolvedMed matches the compact, clean repository structure of `ulamai/UnsolvedMath`: - **Root UnsolvedMath-compatible JSONs:** - `problems.json` — {pub_cnt:,} published open/partially-resolved medical problems (`published = true`) - `dataset.json` — All {tot_cnt:,} canonical records (including 100 calibration/archived/withheld records) - `research_results.json` — Structured literature and status assessments for all {tot_cnt:,} records - `source_registry.json` — {stats['total_registered_sources']:,} provenance sources (papers, guidelines, and `NCT0...` trials) with DOIs, PMIDs, tiers, and license metadata - `statement_audit.json` — Statement quality, rights, and safety review records - `statistics.json`, `categories.json`, `difficulty_levels.json`, `sets.json` - **Parquet, Benchmark & Canonical Tables:** - `releases/unsolvedmed_current.parquet` (v{version}: {pub_cnt:,} published records) & `releases/unsolvedmed_v1_5_current.parquet` (v1.5: 15,000 published records) - `eval/benchmark_v3.jsonl` & `eval/benchmark_summary.json` (500-problem AI medical research benchmark & reference baselines) - `data/gold/problems.jsonl` ({gold_cnt:,} Gold Seed records) - `relations/relations.jsonl` ({stats['total_relations']:,} problem relations) - `history/status_events.jsonl` ({stats['total_status_events']:,} status events) - `schemas/*.schema.json`, `ontology/*.yaml`, `taxonomy/*.yaml`, `prompts/*.md`, `docs/*.md` --- ## 6. Quickstart Usage ```python import json import pyarrow.parquet as pq # Load v{version} (20,000 problems) via Parquet table = pq.read_table("releases/unsolvedmed_current.parquet") print("Published problems in Parquet:", table.num_rows) # Load Gold Seed ({gold_cnt:,} clinician + methodologist verified records) with open("data/gold/problems.jsonl", "r", encoding="utf-8") as f: gold_seed = [json.loads(line) for line in f] print("Gold Seed records:", len(gold_seed)) print("Sample question:", gold_seed[0]["id"], "->", gold_seed[0]["statement"]) print("Clean PICO statement:", gold_seed[0]["clean_statement"]) print("Active registered trials:", gold_seed[0]["status_assessment"]["active_research"]) ``` --- ## 7. Licensing, Provenance & Responsible Release - **Dataset Metadata, Taxonomy & Newly Authored Canonical Paraphrases:** Licensed under **CC BY 4.0** (`LICENSE`). - **Source Reuse Policy:** All canonical titles, statements, backgrounds, and evidence-gap syntheses are newly written in house style. Zero raw documents or verbatim source excerpts are redistributed. - **Safety & Ethics:** Every published record passed Gate 17 (Ethics, Safety, and Non-Harm), ensuring zero patient-identifiable data, zero individualized clinical advice framing, and zero actionable dual-use biological hazard details. """ yaml_frontmatter = f"""--- annotations_creators: - expert-generated - machine-generated language: - en license: cc-by-4.0 multilinguality: - monolingual pretty_name: UnsolvedMedical ({pub_cnt:,} Open Medical, Clinical, Translational, and Public-Health Problems) size_categories: - 10K