|
|
--- |
|
|
license: mit |
|
|
task_categories: |
|
|
- question-answering |
|
|
language: |
|
|
- fr |
|
|
tags: |
|
|
- medical |
|
|
- biology |
|
|
size_categories: |
|
|
- 10K<n<100K |
|
|
--- |
|
|
|
|
|
# MedInjection-FR — Translated Subset 🌍 |
|
|
|
|
|
## Summary |
|
|
|
|
|
The **Translated** component of **MedInjection-FR** adapts large-scale **English biomedical instruction datasets** into French through high-quality automatic translation. |
|
|
It represents the most extensive part of the collection, comprising **416 401 instruction–response pairs**, and provides a bridge between English biomedical resources and French medical instruction tuning. |
|
|
|
|
|
This subset was designed to ensure **broad domain coverage** while maintaining **semantic and linguistic fidelity**, enabling robust cross-lingual transfer and comparative evaluation between native, synthetic, and translated supervision sources. |
|
|
|
|
|
## Motivation |
|
|
|
|
|
The scarcity of large-scale French biomedical instruction data limits the capacity of LLMs to generalize across complex medical domains. |
|
|
To address this, the Translated subset leverages **trusted English benchmarks** spanning medicine, biology, psychology, and clinical knowledge and systematically translates them into French while preserving the structure of instruction–response pairs. |
|
|
|
|
|
This allows for rigorous experiments on **cross-lingual instruction adaptation**, complementing the native and synthetic subsets of MedInjection-FR. |
|
|
|
|
|
## Composition |
|
|
|
|
|
To expand data coverage while maintaining linguistic fidelity, a large collection of **English biomedical instruction datasets** was translated into French using **[Gemini 2.0 Flash](https://blog.google/products/gemini/google-gemini-2/)** and **[GPT-4o-mini](https://openai.com/research/gpt-4o)**. |
|
|
The translated component comprises **416 401 instruction–response pairs** derived from several established English resources: |
|
|
|
|
|
- **[MedQA](https://arxiv.org/abs/2009.13081)** – medical board–style multiple-choice QA. |
|
|
- **[PubMedQA](https://aclanthology.org/D19-1259/)** – factoid biomedical QA from PubMed abstracts. |
|
|
- **[MedMCQA](https://proceedings.mlr.press/v174/pal22a.html)** – clinical multiple-choice QA covering 21 medical specialties. |
|
|
- **[MMLU](https://arxiv.org/abs/2009.03300)** – six medical categories (e.g., anatomy, clinical knowledge, college biology, college medicine, medical genetics, professional medicine). |
|
|
- **[K-QA](https://aclanthology.org/2024.bionlp-1.22/)** – open-ended biomedical question answering for reasoning over scientific text. |
|
|
- **[MMLU-PRO](https://arxiv.org/abs/2406.01574)** – professional-level QA across psychology, biology, and health domains. |
|
|
- **[MedXpertQA](https://arxiv.org/abs/2501.18362)** – clinical reasoning dataset focusing on multi-hop and expert-level diagnostic questions. |
|
|
|
|
|
To assess translation quality, outputs were evaluated using **BLEU** and **COMET** on the **[WMT 2024 Biomedical Translation Task](https://github.com/biomedical-translation-corpora/corpora)** corpus. |
|
|
Both *Gemini 2.0 Flash* and *GPT-4o-mini* achieved results comparable to the **best WMT 2024 system**, indicating that the Translated subset maintains high semantic fidelity and linguistic quality suitable for French biomedical instruction tuning. |
|
|
|
|
|
## Use |
|
|
|
|
|
Intended for: |
|
|
- Fine-tuning French biomedical LLMs with translated instruction–response pairs |
|
|
- Studying cross-lingual transfer and translation quality in instruction tuning |
|
|
- Evaluating the interplay between translation fidelity and domain adaptation performance |
|
|
|
|
|
|
|
|
|
|
|
|
