File size: 3,010 Bytes
102f363
 
 
 
 
 
 
 
 
 
 
7f8203a
 
 
 
 
 
 
67fe52f
7f8203a
 
 
 
 
 
 
 
 
67fe52f
 
87df2c8
 
67fe52f
633512a
5c0e226
67fe52f
633512a
5c0e226
67fe52f
87df2c8
5c0e226
 
633512a
67fe52f
633512a
67fe52f
633512a
67fe52f
 
 
 
 
 
 
 
 
 
 
5c0e226
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
---
license: mit
task_categories:
- question-answering
language:
- fr
tags:
- medical
- biology
size_categories:
- 10K<n<100K
---

# MedInjection-FR — Native Subset 🇫🇷

## Summary

The **Native** component of **MedInjection-FR** comprises **French biomedical instructions and question–answer pairs** natively written in French.  
It forms the core of the dataset and reflects **authentic medical reasoning and linguistic formulations**, sourced from curated corpora and educational materials.  
This subset serves as the **high-quality reference supervision** for instruction tuning of large language models (LLMs) in French biomedical contexts.

## Motivation

Instruction tuning has become essential for adapting LLMs to domain-specific prompts. Yet, in medicine, **native French supervision remains scarce**.  
The Native subset bridges this gap, providing instruction–response data derived directly from **French medical exams, textbooks, and clinical resources**, preserving authentic phrasing, domain structure, and context.

## Composition

The native component combines curated datasets and web-scraped French medical resources to reflect authentic domain knowledge. It integrates the following resources:

- **[S-Editions](https://s-editions.fr/)**
 526 question–answer pairs from a French educational platform for medical students.  

- **[MediQAl](https://github.com/abazoge/MediQAl)**:  
  32 603 items from national medical examinations covering 41 medical specialties.  

- **[FrenchMedMCQA](https://arxiv.org/abs/2304.04280)**:  
  3 105 pharmacy-focused multiple-choice questions.  

- **[mlabonne/medical-cases-fr](https://huggingface.co/datasets/mlabonne/medical-cases-fr)** and **[mlabonne/medical-mcqa-fr](https://huggingface.co/datasets/mlabonne/medical-mcqa-fr)**:  
  12 194 examples originating from French medical exam databases.  

- **[FrBMedQA](https://www.researchgate.net/publication/365908538_FrBMedQA_the_first_French_biomedical_question_answering_dataset)**:  
  19 836 questions derived from French biomedical Wikipedia articles spanning eight UMLS semantic groups (chemicals and drugs, anatomy, physiology, disorders, phenomena, procedures, genes and molecular sequences, and devices).  
  Originally closed-form, these questions were reformulated into multiple-choice format using *[GPT-4o-mini](https://openai.com/research/gpt-4o)* for standardization.  

- **[Biomedical Translation Corpora (WMT)](https://github.com/biomedical-translation-corpora/corpora)**:  
  Bilingual biomedical translation data from the WMT challenge repositories were reformulated into instruction–response pairs.  
  Each instruction requests the **French translation** of an English biomedical passage, reframing translation as an instruction-following task aligned with the native portion.

## Use

Used to fine-tune biomedical LLMs for:
- Question answering and clinical reasoning  
- Instruction-following in French  
- Evaluating cross-domain instruction generalization