README.md

---
pretty_name: MamaRetrieval
language:
  - en
license: other
license_name: mamaretrieval-research-only-v1
license_link: LICENSE
task_categories:
  - text-retrieval
  - question-answering
tags:
  - medical
  - clinical-guidelines
  - midwifery
  - obstetrics
  - retrieval-benchmark
  - evaluation
  - llm-as-judge
  - rag
size_categories:
  - 1K<n<10K
configs:
  - config_name: queries
    data_files:
      - split: test
        path: "data/queries.parquet"
  - config_name: rankings
    data_files:
      - split: test
        path: "data/rankings.parquet"
  - config_name: judgments
    data_files:
      - split: test
        path: "data/judgments.parquet"
  - config_name: chunks
    data_files:
      - split: test
        path: "data/chunks.parquet"
  - config_name: judgments_with_reasoning
    data_files:
      - split: test
        path: "audit/judgments_with_reasoning.parquet"
---

# MamaRetrieval — v0.1.0

A retrieval evaluation benchmark for medical RAG systems serving midwives and
doctors. 3,185 clinical queries on midwifery / OBGYN topics, evaluated against
the top-3 results of 6 retrievers, with per `(query, chunk)` pair labels graded
by an LLM judge under a four-dimension rubric.

This release is the **Tier 2** split (top-3 union of 6 retrievers, 36,418
labelled `(q, c)` pairs). The Tier 3 split (top-20 union) will land as v0.2.0.

## Quick start

```python
from datasets import load_dataset

queries   = load_dataset("nmrenyi/mamaretrieval", "queries",   split="test")
rankings  = load_dataset("nmrenyi/mamaretrieval", "rankings",  split="test")
judgments = load_dataset("nmrenyi/mamaretrieval", "judgments", split="test")
chunks    = load_dataset("nmrenyi/mamaretrieval", "chunks",    split="test")

# Optional — the same judgments + the judge's per-row reasoning trace (~117 MB)
judgments_full = load_dataset("nmrenyi/mamaretrieval",
                              "judgments_with_reasoning", split="test")
```

## Configs

| Config | Rows | Columns | What it is |
|---|---:|---|---|
| `queries` | 3,185 | `query_id`, `query_text`, `seed_chunk_id` | The benchmark queries, each generated by an LLM from a single chunk of the corpus. |
| `rankings` | 57,330 | `query_id`, `retriever`, `rank`, `chunk_id`, `score` | For every query × retriever combination, the top-3 `chunk_id`s with the retriever's similarity score. 6 retrievers × 3,185 queries × 3 = 57,330. |
| `judgments` | 36,418 | `query_id`, `chunk_id`, `d1_topic`, `d2_meaningful`, `d3_actionable`, `d4_density`, `score` | One label per unique `(query, chunk)` pair in the pooled top-3 union. `score = d1 × (d2 + d3 + d4) ∈ [0..6]`. |
| `judgments_with_reasoning` | 36,418 | (same as `judgments`) + `thinking` | The same labels with the judge model's reasoning trace per row. Ships in `audit/` because it's ~117 MB and not needed to use the benchmark. |
| `chunks` | 17,827 | `chunk_id`, `text` | The chunk text for every `chunk_id` referenced by `queries.seed_chunk_id` or any retriever's top-3 result. Drawn from the producer corpus (see Provenance). |

### Schema notes

- **`chunk_id`** is the 16-character hexadecimal identifier from the
  producer corpus. Every `chunk_id` that appears in `rankings`, `judgments`,
  `judgments_with_reasoning`, or `queries.seed_chunk_id` is guaranteed to be
  resolvable in `chunks`.
- **`score`** in `judgments` is computed downstream from the four dimensions
  via `score = d1 × (d2 + d3 + d4)`. The judge emits only `d1..d4`.
- **`seed_chunk_id`** records which chunk an LLM was given when it
  synthesised the query. It's provenance, **not** a gold label — seed
  chunks may not appear in any retriever's top-3, and when they do they
  are not always the highest-rated chunk for that query.

## Rubric

The judge scores each `(query, chunk)` pair on four dimensions:

- **D1 — Topic** (bool): does the chunk address the same clinical problem as
  the query (same condition, intervention, *and* clinical-timing context)?
  If `D1 = false`, `D2 = D3 = D4 = 0` automatically.
- **D2 — Meaningful clinical content** (0–2): how rich is the chunk's
  clinical content, independent of whether it specifically answers the query?
- **D3 — Actionable guidance** (0–2): how specific is the actionable
  guidance — vague advice (0), general direction (1), exact
  doses/thresholds/steps (2)?
- **D4 — Density** (0–2): what fraction of the chunk is directly useful for
  answering *this specific* query?

`score = d1 × (d2 + d3 + d4) ∈ [0..6]`.

The full prompt — including four worked examples that anchor the
calibration — is shipped verbatim at
[`audit/judge_relevance_prompt.txt`](audit/judge_relevance_prompt.txt). Its
`prompt_hash` is recorded in [`manifest.json`](manifest.json).

## Retrievers

| `name` | `model` |
|---|---|
| `bm25` | BM25 (lexical baseline) |
| `medcpt` | `ncbi/MedCPT` (Query + Article encoders) |
| `octen` | `Octen/Octen-Embedding-8B` |
| `voyage` | `voyage-4-large` |
| `lateon` | `lightonai/GTE-ModernColBERT-v1` (late-interaction ColBERT) |
| `gecko` | `gecko-1024-quant-v0.2.0` (on-device TFLite, deployed retriever) |

All retrievers were run on the producer corpus (see Provenance) and their
top-20 results stored. This release exposes the **top-3** of each — the
deployment-honest depth for the RAG system this benchmark was built for.

## How the dataset was made

1. **Query generation.** For each clinically-relevant chunk in the producer
   corpus, an LLM (`Qwen/Qwen3.6-27B-FP8`) was prompted to produce one
   ≤20-word clinical question the chunk could answer. Chunks judged
   non-clinical (e.g. course outlines, references, learning objectives)
   were skipped. The full prompt is shipped at
   [`audit/query_generation_prompt.txt`](audit/query_generation_prompt.txt).
2. **Retrieval.** Each query was run against the producer corpus by every
   retriever. Top-20 candidates per retriever were stored.
3. **Pooling.** For each query, the union of every retriever's top-3 was
   deduped (~11.4 unique chunks per query at this scale).
4. **Judging.** Every `(query, chunk)` pair in the pool was scored by
   `Qwen/Qwen3.5-397B-A17B-FP8` against the four-dimension rubric. The
   judge's reasoning was captured separately and is shipped in
   `judgments_with_reasoning`.

Validation: the judge model was calibrated against Claude Opus 4.7
reference labels on a 62-pair pilot, with 95% threshold agreement at score ≥
3 and 85% at score ≥ 5.

## Provenance

- **Producer corpus**: `rag-bundle-v0.2.0`, produced at commit
  [`a1abe003`](https://github.com/nmrenyi/mamai-medical-guidelines/tree/a1abe003cce742b46954375d17abb28a3e27110f)
  of [`nmrenyi/mamai-medical-guidelines`](https://github.com/nmrenyi/mamai-medical-guidelines).
  The 63,650-chunk corpus the retrievers were run against. Built from a mix
  of WHO guidelines, Tanzania / Zanzibar MOH documents, and a small set of
  midwifery references.
- **Versioning**: `v0.1.0` = Tier 2 (top-3 union). `v0.2.0` will add Tier 3
  (top-20 union) on the same query set when judging finishes.
- **Audit trail**: [`manifest.json`](manifest.json) pins exact judge and
  generator model IDs, prompt hashes, and schema versions.

## License — **Research use only**

This dataset is released for **non-commercial academic research and
retrieval-evaluation benchmarking only**. By downloading or using it, you
agree to all of the following:

**Permitted**

- Academic research, including publication of aggregate metrics, qualitative
  analysis, ablations, and methodology comparisons.
- Use as an *evaluation* benchmark for retrieval systems.
- Re-running the rubric or running new judges against the included
  `(query, chunk)` pairs for methodology research.

**Not permitted without explicit written permission**

- Any commercial use, including evaluation as part of internal product
  decisions at for-profit organisations.
- Use of the chunk text as **training data** for any model — generative,
  embedding, retrieval, or otherwise.
- Redistribution of the chunk text, in whole or in part, outside the form
  shipped here (i.e. do not extract `chunks.parquet`, repackage, mirror, or
  re-host the chunk content).
- Production deployment of any system whose retrieval or judging behaviour
  has been tuned on this data.
- Clinical use of the chunk text. **None of the chunk content has been
  reviewed for clinical accuracy in the form presented here**; do not surface
  it to patients or clinicians.

Full terms — including upstream-licensing constraints, attribution, and
warranty disclaimers — are in [LICENSE](LICENSE).

## Citation

Ren, Yi. *MamaRetrieval* v0.1.0. 2026. <https://huggingface.co/datasets/nmrenyi/mamaretrieval>

## Limitations

- **Scope**: midwifery / OBGYN / neonatal care, framed for guidelines deployed
  in Zanzibar. Performance numbers do not transfer cleanly to general
  medical retrieval.
- **Depth-3 ceiling**: ~25% of queries have no `score ≥ 5` chunk in any
  retriever's top-3, even from the strongest retriever. This is an inherent
  depth-3 pool limit, not a retriever failure.
- **Single relevance judge**: every `(query, chunk)` relevance label in
  this dataset is produced by one LLM (`Qwen/Qwen3.5-397B-A17B-FP8`) under
  the four-dimension rubric. That judge was calibrated against Claude Opus
  4.7 on a 62-pair pilot — 95% threshold agreement at score ≥ 3, 85% at
  ≥ 5 — but that's a small LLM-vs-LLM sanity check, not a human-annotated
  gold standard. Practical consequences: retriever-vs-retriever rankings
  tend to be stable across reasonable relevance judges, but absolute
  score distributions and per-row labels will shift if you re-grade the
  same `(query, chunk)` pairs with a different judge. Treat each label as
  one judge's calibrated opinion, not ground truth.