README.md

---
license: cc-by-4.0
language:
  - en
pretty_name: Cancer Knowledge Base
tags:
  - medical
  - oncology
  - cancer
  - knowledge-base
  - clinical-education
  - rag
  - retrieval-augmented-generation
  - benchmark
  - evaluation
  - llm
  - ehr
  - multiple-choice
  - icd-10
  - loinc
  - icd-o-3
  - hgnc
  - localization
  - multilingual
  - seer
  - guidelines
  - nccn
  - esmo
  - asco
  - synthetic-data
  - text2sql
size_categories:
  - n<1K
task_categories:
  - question-answering
  - text-generation
  - text-retrieval
task_ids:
  - open-domain-qa
  - closed-domain-qa
  - multiple-choice-qa
  - named-entity-recognition
---
# Cancer Knowledge Base — the open, verified oncology KB for RAG & LLM evaluation

The only open **CC-BY-4.0** oncology knowledge base that combines:

- **110/110 trials cited** with PMID + NCT + PubMed/ClinicalTrials.gov URLs, and **32 prognosis
  rows linked to verified SEER 2016–2022 references** — no LLM-synthetic dataset has this.
- **A provable 152-question MCQ benchmark** — every answer derives from this KB's own structured
  data and carries a citation + golden docs, so it is *open-book verifiable* (unlike exam-based
  medical benchmarks).
- **An oncology RAG evaluation suite** — 1,239-chunk retrieval pool, 456 MedRGB-style
  sufficiency / noise / counterfactual-robustness variants, MIRAGE/MedRGB interchange exports,
  a dependency-free harness (`eval_mcq.py`), and a submission package.
- **A ground-truth validation layer** that independently audits LLM-synthetic oncology datasets
  against the KB's cited statements — published scorecards for CancerGUIDE, OncoBench, and
  MedGUIDE-MCQA-8K.
- **Universal TNM / staging + sub-stages** (65 staging rows) and **ICD-10 / ICD-O-3 / LOINC / HGNC
  interop codes** for joins into precision-oncology pipelines.
- **Localization with provenance**: 9 locales, 4,728 glossary terms, ≥60% KB-wide translatable
  coverage for every language, and 20 reviewed-as-auto translation rows.

Educational reference only — **not medical advice**. Fully de-identified by construction.

## Quick start

```python
import pandas as pd

documents = pd.read_parquet("hf://datasets/ranjithraj/cancer-knowledge-base/documents.parquet")
guidelines = pd.read_parquet("hf://datasets/ranjithraj/cancer-knowledge-base/guidelines.parquet")
trials = pd.read_parquet("hf://datasets/ranjithraj/cancer-knowledge-base/evidence_levels.parquet")

# Full-text search any document (FTS is a first-class column)
breast = documents[documents.path == "know/breast-ref.md"].body.iloc[0]

# Every guideline statement, citation-linked
print(guidelines[guidelines.source == "know/breast-ref.md"])

# 152 provable MCQs with difficulty / evidence / citation / golden docs
mcq = pd.read_parquet("hf://datasets/ranjithraj/cancer-knowledge-base/mcq_questions.parquet")
```

Evaluate your own RAG/LLM on the oncology benchmark in one command:
```bash
python3 scripts/eval_mcq.py --retriever bm25 --topk 3   # BM25 end-to-end
python3 scripts/eval_mcq.py --predictions preds.jsonl   # score your system
```

A 3-minute Colab walkthrough is at [`colab_quickstart.ipynb`](colab_quickstart.ipynb) — load the KB,
run BM25 retrieval, and reproduce the benchmark baseline.

## Suggested uses

- **RAG evaluation** — score your retriever+generator on `benchmark/` (MIRAGE/MedRGB interchangeable).
- **Fine-tuning** — instruction-tune or domain-adapt small LLMs for oncology QA on the 58 de-identified documents.
- **Benchmarking synthetic data** — screen LLM-generated oncology datasets against the KB's evidence index (`scripts/validate_synthetic.py`).
- **Multilingual oncology content** — build localized Q&A or terminology tools from the code-anchored glossary (9 locales).
- **Entity linking / coding** — ICD-10 / LOINC / ICD-O-3 / HGNC mappings in `codes.parquet`.

## Citation

```bibtex
@misc{cancer-knowledge-base-2026,
  title = {{Cancer Knowledge Base}: an open, verified, provable oncology knowledge base for RAG and LLM evaluation},
  author = {Ranjithraj, {R}},
  year = {2026},
  howpublished = {\url{https://huggingface.co/datasets/ranjithraj/cancer-knowledge-base}},
  note = {CC-BY-4.0; 110/110 trials cited; 152-question provable MCQ benchmark; validation scorecards for CancerGUIDE/OncoBench/MedGUIDE}
}
```

## Contents (parquet files)

| File | Rows | Description |
|------|------|-------------|
| `documents.parquet` | 59 | De-identified documents: metadata + body text |
| `sections.parquet` | 436 | Markdown headings + verbatim section content |
| `markdown_tables.parquet` | 225 | All markdown tables (header/rows as JSON) |
| `markdown_links.parquet` | 168 | Internal links between documents |
| `prognosis.parquet` | 168 | Lethality/survival tables by cancer type and stage (dated, educational) |
| `biomarkers.parquet` | 72 | Mutations, receptors, genetic syndromes |
| `guidelines.parquet` | 175 | Guideline-grade treatment statements per cancer (NCCN/ESMO/ASCO) |
| `subtypes.parquet` | 110 | Histologic subtypes by cancer |
| `drugs.parquet` | 141 | Drugs, class, mechanism, toxicity |
| `staging.parquet` | 65 | Staging systems + universal TNM/Stage 0-IV + sub-stages |
| `regimens.parquet` | 85 | Treatment regimens per case and phase |
| `monitoring.parquet` | 109 | Factors to monitor during chemo/targeted/surgery/RT |
| `diagnostics.parquet` | 96 | Recommended diagnostics per case |
| `labs.parquet` | 72 | Baseline labs (synthetic, clearly flagged) |
| `imaging_findings.parquet` | 59 | PET-CT / imaging findings (synthetic, clearly flagged) |
| `consultations.parquet` | 59 | Consultation plan evolution (synthetic, clearly flagged) |
| `case_facts.parquet` | 218 | Structured case fields (diagnosis, staging, prognosis) |
| `evidence_levels.parquet` | 115 | Evidence grading (guidelines + trials) per entity — 110 trials with PMID/NCT/URL |
| `codes.parquet` | 82 | Interop codes: ICD-10 sites/types, HGNC biomarkers, LOINC labs, ICD-O-3 breast subtypes |
| `regimen_drugs.parquet` | 100 | Regimen-to-drug join table |
| `mcq_questions.parquet` | 152 | 152 provable oncology MCQs with difficulty / evidence / citation / golden docs |
| `prognosis_sources.parquet` | 32 | Prognosis rows linked to verified SEER references (vintage + value) |
| `locales.parquet` | 9 | Registered target languages (BCP-47 codes, review requirement, priority) |
| `glossary.parquet` | 4728 | Code-anchored terminology registry: concept -> code (ICD/LOINC/HGNC/ICD-O-3) -> local term per locale |
| `translations.parquet` | 20 | Per-document localized bodies with master_hash/content_hash provenance and review status |

## RAG Benchmark

`benchmark/` contains an oncology RAG evaluation suite built from this knowledge base:
- `retrieval_pool.parquet` — 1,239 chunked retrieval units (sections, markdown tables, domain tables)
- `mcq_benchmark.parquet` — 152 multiple-choice questions with golden docs, difficulty, evidence level, and citations
- `mcq_robustness.parquet` — 456 MedRGB-style variants (sufficiency / noise / counterfactual-robustness)
- `mirage_format.jsonl` / `medrgb_format.jsonl` — interchange exports for MIRAGE-family leaderboards and MedRGB harnesses (see `benchmark/MIRAGE_INTEGRATION.md`)
- `leaderboard.json` + `releases/*.json` — recorded results with KB-commit / date / release snapshots for growth tracking
- `baseline_preds.jsonl` + `SUBMISSION.md` — reproducible baseline and submission protocol

Baseline (BM25@top3, no LLM): **sufficiency 31.6%** (random = 25%), **retrieval@3 61.8%**. Real
headroom for embedding retrievers and strong generators — the pool is provably answerable, so
gains are attributable to retrieval + reasoning, not answer leakage.

## Ground-truth validation scorecards

`scorecards/` holds the results of screening LLM-synthetic oncology datasets against the KB's cited
statements (`scripts/validate_synthetic.py`). Each claim is tiered **supported / partial / unsupported**
by lexical coverage of site-gated evidence, and every item links to the matching evidence chunk.

- `cancerguide.all.summary.md` — CancerGUIDE (316 synthetic patient notes): 238 scorable claims,
  **87.0% supported** / 12.6% partial / 0.4% unsupported.
- `oncobench.all.summary.md` — OncoBench strong100 (100 adjudicated decision cases): 2,007 claims,
  **26.0% supported** — the scorecard doubles as a KB coverage-gap map.
- `medguide.all.summary.md` — MedGUIDE-MCQA-8K (7,747 NCCN decision-tree MCQs): 6,952 claims,
  **21.2% supported**.

## Localization (l10n)

The KB is localized *with provenance* — every translation keeps its citation (via the master
document), a `master_hash` so a change in the English source flags the locale stale, and an explicit
review status (`auto` → `reviewed` → `approved`). **All 9 registered locales are at ≥60% KB-wide
translatable coverage** (de/es/hi/zh ~520-525 terms, ar/bn/pt-BR/sw/te 527 terms each; 4,728
glossary terms total), measured script-agnostically across all 58 documents.

- `locales.parquet` — registered languages (de, zh, es, hi, ar, bn, pt-BR, sw, te)
- `glossary.parquet` — **code-anchored terminology registry**: each concept is linked to its
  language-neutral code (ICD-10 / ICD-O-3 / LOINC / HGNC) and maps to the local clinical term,
  so structured rows localize for free while codes stay the join key across languages.
- `translations.parquet` — localized document bodies (5 reference docs at 100% translatable
  coverage in de/zh/es/hi) with `master_hash`/`content_hash` and `status`.
- `l10n/*/mcq.*.jsonl` — localized MCQ exports (de/es/hi/zh) with preserved answer indices.

**Coverage is not clinical approval**: `status='auto'` rows are glossary-generated drafts that must
be clinically reviewed (`reviewed`) before use. See `l10n-moat.md` and
`scripts/localize_pilot.py --coverage`.

## Privacy

This is the **de-identified public subset**. The real patient case (`cases/specific/` — PHI) and all 
content derived from it are excluded. Documents are filtered by `phi = 0`.

Synthetic rows in `labs`, `imaging_findings`, and `consultations` are **clearly flagged** with 
`synthetic = 1` and are illustrative public-case data — they are not real patient values.

## Disclaimer

Educational reference only — **not medical advice**. Survival figures are population averages and may 
be outdated (`prognosis.as_of` marks their review date); consult current guidelines and clinicians for 
patient decisions.