benchmark/README.md

---
license: cc-by-4.0
language:
  - en
tags:
  - medical
  - oncology
  - rag
  - benchmark
  - evaluation
  - llm
  - retrieval-augmented-generation
  - mirlig
size_categories:
  - n<1K
---

# Oncology RAG Benchmark — provable, MIRAGE/MedRGB-ready, open

An oncology RAG evaluation suite whose answers are **provable from the KB's own
structured data** — every question carries a citation (PMID / NCT / SEER URL)
and `golden_docs` (the retrieval units that contain the answer).

- **152 questions** across prognosis / staging / subtypes / drugs / codes /
  biomarkers / prevention, each with `difficulty`, `evidence_level`,
  `citation`, and `golden_docs`.
- **456 MedRGB-style variants** — sufficiency (golden docs only), noise
  (golden + 5 irrelevant docs), robustness (golden + a counterfactually edited
  doc). The robustness rows carry an explicit `counterfactual_doc`.
- **1,239-chunk retrieval pool** (sections, markdown tables, domain tables) for
  end-to-end RAG evaluation.
- **MIRAGE + MedRGB interchange exports** (`mirage_format.jsonl`,
  `medrgb_format.jsonl`) for ingestion by external harnesses/leaderboards.
- **Dependency-free harness** (`eval_mcq.py`: BM25 / embeddings / predictions),
  a reproducible baseline (`baseline_preds.jsonl`), and a submission protocol
  (`SUBMISSION.md`).

## Baseline (BM25@top3, no LLM)

| Scenario | Accuracy | vs random (25%) |
|----------|----------|------------------|
| sufficiency | 31.6% | +6.6 pts |
| noise | 31.6% | +6.6 pts |
| robustness | 31.6% | +6.6 pts |
| retrieval@3 | 61.8% | — |

Real headroom for embedding retrievers and strong generators. Because every
answer is provable, score differences reflect genuine retrieval + reasoning,
not answer leakage.

## Reproduce

```bash
# BM25 end-to-end (retrieval + answer)
python3 eval_mcq.py --retriever bm25 --topk 3

# Score your own system's predictions ({"question_id": 1, "answer_index": 3})
python3 eval_mcq.py --predictions preds.jsonl

# Verify the shipped baseline
python3 eval_mcq.py --predictions baseline_preds.jsonl
```

## Citation

```bibtex
@misc{oncology-rag-benchmark-2026,
  title = {{Oncology RAG Benchmark}: a provable, MIRAGE/MedRGB-ready oncology evaluation suite},
  author = {Ranjithraj, {R}},
  year = {2026},
  howpublished = {\url{https://huggingface.co/datasets/ranjithraj/cancer-knowledge-base}},
  note = {CC-BY-4.0; 152 questions; 456 MedRGB variants; answers provable from cited KB data}
}
```

## Files

| File | Contents |
|------|----------|
| `mcq_benchmark.parquet` | 152 questions (question, choices, answer_key, answer, difficulty, evidence_level, citation, golden_docs) |
| `mcq_robustness.parquet` | 456 variants (sufficiency / noise / robustness + counterfactual_doc) |
| `retrieval_pool.parquet` | 1,239 retrieval units (sections, tables, domain tables) |
| `mirage_format.jsonl` | 152 MIRAGE-compatible question instances |
| `medrgb_format.jsonl` | 456 MedRGB variants (sufficiency / noise / robustness) |
| `leaderboard.json` + `releases/` | recorded systems with KB-commit / date / release snapshots |
| `baseline_preds.jsonl` | reproducible BM25@top3 predictions |
| `SUBMISSION.md` | reference-benchmark submission protocol |
| `MIRAGE_INTEGRATION.md` | MIRAGE/MedRGB integration guide |

## What makes it *provable*

Unlike exam-style medical benchmarks (MedQA, PubMedQA, MedXpertQA), each answer
here is a row in the source KB with a citation chain. A grader can audit any
answer against the cited chunk — the benchmark is *open-book verifiable*.

## Notes

Built from the [Cancer Knowledge Base](https://huggingface.co/datasets/ranjithraj/cancer-knowledge-base)
(CC-BY-4.0). Educational reference only — **not medical advice**.