---
title: fully-open-meditron
description: "The training corpus of Fully Open Meditron (EPFL, LiGHT laboratory): about 601k examples (~150M tokens) for supervised fine-tuning of clinical LLMs. Eight public medical QA datasets, plus three clinician-vetted synthetic parts: exam-style QA, QA grounded in 46,469 clinical practice guidelines, and open clinical vignettes. Decontaminated against the evaluation sets."
tags:
- clinical
- healthcare
- llm
- medical
- meditron
- qa
- sft
license: apache-2.0
source_url: https://huggingface.co/datasets/EPFLiGHT/fully-open-meditron
---

# fully-open-meditron

> **Mirrored metadata.** This card is reproduced on Atlas for discovery. The artifact itself is not hosted here — follow the source link for the files, and refer to the upstream licence for terms of use.

The training corpus of Fully Open Meditron (EPFL, LiGHT laboratory): about 601k examples (~150M tokens) for supervised fine-tuning of clinical LLMs. Eight public medical QA datasets, plus three clinician-vetted synthetic parts: exam-style QA, QA grounded in 46,469 clinical practice guidelines, and open clinical vignettes. Decontaminated against the evaluation sets.

Upstream identifier: `EPFLiGHT/fully-open-meditron`. Paper: [arXiv:2605.16215](https://arxiv.org/abs/2605.16215).

**Source:** https://huggingface.co/datasets/EPFLiGHT/fully-open-meditron

**Licence:** apache-2.0

_This dataset is part of the Atlas demo catalogue. See the repository history for provenance._