datasets/meditron/data/README.md

---
license: cc-by-nc-4.0
language:
  - en
task_categories:
  - question-answering
  - text-generation
tags:
  - medical
  - clinical
  - healthcare
  - llm
  - sft
size_categories:
  - 100K<n<1M
pretty_name: Fully Open Meditron Corpus
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/*/train-*.parquet
  - config_name: curated_qa
    data_files:
      - split: train
        path: data/curated_qa/train-*.parquet
  - config_name: synthetic_curated_qa
    data_files:
      - split: train
        path: data/synthetic_curated_qa/train-*.parquet
  - config_name: guidelines_qa
    data_files:
      - split: train
        path: data/guidelines_qa/train-*.parquet
  - config_name: synthetic_moove
    data_files:
      - split: train
        path: data/synthetic_moove/train-*.parquet
---

# Fully Open Meditron Corpus

A clinician-vetted training corpus for medical large language models, accompanying the paper *Fully Open Meditron: An Auditable Pipeline for Clinical LLMs* (anonymous submission to NeurIPS 2026 Evaluations & Datasets Track).

The corpus combines eight aggregated public medical QA datasets with three clinician-vetted synthetic components, totaling approximately 601k examples (~150M tokens). It is designed to support supervised fine-tuning of large language models for clinical decision support and medical question answering, with full transparency over data provenance, generation prompts, and decontamination.

## Quick Start

```python
from datasets import load_dataset

# Load the full merged corpus (default)
ds = load_dataset("meditron-fo-anon/fully-open-meditron")

# Load a single component (e.g. for ablations)
ds = load_dataset("meditron-fo-anon/fully-open-meditron", "synthetic_moove")
```

## Components

| Config | Examples | Description |
|---|---|---|
| `curated_qa` | 216,546 | Aggregated public medical QA training splits (MedQA, MedMCQA, PubMedQA, MedExpQA, HealthSearchQA, LiveQA, AfriMed-QA v1/v2), normalized into (system, user, assistant) conversational format. 173 items removed by system-wide decontamination. |
| `synthetic_curated_qa` | 214,654 | Novel exam-style QA generated by gpt-oss-120b, seeded from the curated pool, stratified by question type with continuous answer-position monitoring to prevent label bias. |
| `guidelines_qa` | 145,681 | QA grounded in 46,469 clinical practice guidelines from 16 global institutions. |
| `synthetic_moove` | 24,465 | Open-ended clinical vignette prompts seeded from an expert-written vignette pool, designed to elicit complex diagnostic reasoning. |
| **Total** | **601,346** | |

The `default` config concatenates all four.

## Schema

| Field | Type | Description |
|---|---|---|
| `id` | string | Unique identifier |
| `messages` | list of `{role, content}` | Conversation in OpenAI-style format. Roles: `system`, `user`, `assistant`. |
| `source_component` | string | One of `curated_qa`, `synthetic_curated_qa`, `guidelines_qa`, `synthetic_moove`. |
| `is_synthetic` | bool | Whether the row was generated by an LLM teacher. |
| `teacher_model` | string | Teacher model used for generation (`gpt-oss-120b` or null for source items). |
| `source_dataset` | string | Original public dataset name (curated_qa rows only). |
| `gold_label` | string | Multiple-choice gold answer letter, where applicable. |
| `label_text` | string | Multiple-choice gold answer text, where applicable. |
| `exact_match` | bool | Whether teacher prediction matched the gold label after rejection-sampling resampling. |
| `try_count` | int | Number of resampling attempts required (1–8). |

## Construction

The corpus was constructed in three stages:

1. **Aggregation.** Eight public medical QA datasets were normalized into a unified conversational schema. Items that could not be unambiguously mapped were discarded.
2. **Clinician-vetted synthetic generation.** A four-physician panel reviewed three sampled outputs per few-shot generation prompt template, with disagreements resolved by panel discussion. The audit produced four structural changes to the generation pipeline: tightening overbroad constraints on "controversial" and "outdated" content; requiring explicit disease progression and geographic context; decoupling stems from answers; and excluding overly US-centric phrasing. Synthetic components were then generated by gpt-oss-120b.
3. **Hallucination mitigation.** For every multiple-choice item carrying a labeled answer, the predicted letter was extracted via dataset-specific regex and resampled independently up to 8 times at temperature 0.7 until the extracted letter matched the gold label.

## Decontamination

System-wide two-stage n-gram and token-alignment decontamination was applied against all evaluation references used in the accompanying paper: MedQA, MedMCQA, PubMedQA, MedXpertQA test splits, an open-ended clinical evaluation split, HealthBench Hard, MMLU-Pro, IFEval, and ARC-Challenge. Decontamination is syntactic rather than semantic.

## Licensing

The synthetic components (`synthetic_curated_qa`, `guidelines_qa`, `synthetic_moove`) are released under CC BY-NC 4.0 for research use.

The `curated_qa` component is a derived aggregation of publicly released datasets, each retaining its original license. Users redistributing this component should consult the original licenses of MedQA, MedMCQA, PubMedQA, MedExpQA, HealthSearchQA, LiveQA, and AfriMed-QA v1/v2.

## Limitations

- Approximately 64% of items (71% of tokens) are synthetic, generated by a single teacher (gpt-oss-120b), which introduces model-specific stylistic and reasoning biases.
- Clinician audit covered three sampled QA pairs per generation prompt template, bounding systematic but not item-level errors.
- Decontamination is syntactic (n-gram and token alignment) rather than semantic, leaving open the possibility of paraphrased leakage from teacher generations.
- Coverage of non-English clinical content is limited.
- Inherits geographic and demographic biases of the source datasets (predominantly North American and European clinical contexts), partially mitigated by AfriMed-QA v1/v2.

## Intended Use

For research, including reproducibility, auditing, and red-teaming of medical LLMs. **Not intended as a substitute for clinical judgment.** Models trained on this corpus should not be deployed without independent domain-specific safety evaluation.

## Citation

```bibtex
@inproceedings{anonymous2026meditron,
  title={Fully Open Meditron: An Auditable Pipeline for Clinical LLMs},
  author={Anonymous},
  booktitle={NeurIPS 2026 Evaluations and Datasets Track},
  year={2026},
  note={Under review}
}
```