HealthDataHub/PARCOMED_research_only

PARCOMED - PARTAGES Corpus of Open MEdical Documents

This document describes the first version of the research-only corpus.

Overview

The availability of French biomedical data remains a major challenge for improving the multilingual capabilities of large language models (LLMs) in the medical domain. We introduce and release the PARCOMED_research_only corpus, a collection of French biomedical texts compiled from a wide range of sources for research-only use.

While similar datasets have been released in the past couple of years (NACHOS from DrBERT, JARGON), ours is the result of a greater scrutiny of the licensing terms of each source. Therefore, the PARTAGES corpus is fully compatible with research usage and is also distributed with a version compatible with commercial usage. Here, we present the research-only corpus released.

Document types and data sources

The selected datasets for our corpus come from a variety of sources which can be categorized as follows:

Clinical

E3C: E3C corpus of clinical cases in French, used for training and evaluating medical models. License 'libre for research'.

CAS: Corpus built from clinical cases reported in the scientific literature published in French, of which a subset of the corpus is annotated. NACHOS versioning. Visible at https://huggingface.co/datasets/bigbio/cas/tree/main and available upon request to the author. Research-only license.

FRASIMED: Annotated corpus of synthetic clinical cases written in French. Available at https://zenodo.org/records/8355629. License CC-BY-4.0.

ESSAI: Dataset ESSAI containing annotations of medical texts in French. Not available online but possible upon request. Research-only license.

Dialogue

PXCORPUS: French corpus of medical dialogues on prescriptions, transcripted and annotated. Available at https://doi.org/10.5281/zenodo.6482586. License CC-BY-4.0.

MQC: Annotated corpus of medical dialogues in French, simulating consultations between doctor and patient. Available at https://github.com/kleag/labforsims2-corpus. License CC-BY-SA-NC-4.0.

Education

CERIMES: Indexing of digital pedagogical resources proposed by higher education institutions and research organizations in France. NACHOS versioning. Available at https://data.enseignementsup-recherche.gouv.fr/explore/dataset/fr_esr_ressources-pedagogiques/export/?flg=en-gb&refine.lom_lifecycle_contribute_entity_fn=CERIMES. License Etalab.

Encyclopedic

WIKIPEDIA: Corpus extracted from Wikipedia in French, collected via the python wikipediaapi on medical, pharmaceutical and biological categories. License CC-BY-SA 3.0, GNU Free Documentation License.

Medical

ECDC_TM: Corpus of medical texts from the European Centre for Disease Prevention and Control (ECDC) for machine translation tasks. NACHOS versioning. Available at https://joint-research-centre.ec.europa.eu/language-technology-resources/ecdc-translation-memory_en#Introduction. Free License.

Medicinal

EMEA_V3: Corpus of multilingual medical documents from the European Medicines Agency (EMEA), 3rd version. NACHOS versioning. Available at https://huggingface.co/datasets/qanastek/EMEA-V3. License CC-BY-4.0.

BDPM: Public database of medicines. NACHOS versioning. Available at https://www.data.gouv.fr/fr/datasets/base-de-donnees-publique-des-medicaments-base-officielle/. License Etalab.

Question Answering

DEFT2021: Corpus from the DEFT challenge for three tasks: extraction of clinical profiles, evaluation of student responses and existing ratings. Available at https://huggingface.co/datasets/DrBenchmark/DEFT2021. License CC-BY-4.0.

FRENCHMEDMCQA (INSTRUCT): Francophone corpus of questions in the medical domain with 5 response options (single or multiple choice) and their manual corrections. Available at https://huggingface.co/datasets/qanastek/frenchmedmcqa. License Apache 2.0.

MEDIQAL (INSTRUCT): MediQAl is a French medical question answering dataset designed to evaluate the capabilities of language models in factual medical recall and clinical reasoning. Disponible à https://huggingface.co/datasets/ANR-MALADES/MediQAl. Licence CC-BY-4.0

Regulation

QUALISCOPE: Data on the quality of healthcare establishments in France, extracted from Scope Santé. NACHOS versioning. Available at https://www.data.gouv.fr/fr/datasets/base-sur-la-qualite-et-la-securite-des-soins-anciennement-scope-sante/. License Etalab.

CNEDIMTS: Dataset from a specialized commission of the HAS that evaluates individual medical devices as well as diagnostic, therapeutic or assistive products (excluding medications), as well as associated services. NACHOS versioning. Available at https://www.data.gouv.fr/datasets/evaluation-des-dispositifs-medicaux/. License Etalab.

Scientific

WMT16: Biomedical variant of the WMT16 corpus built from PubMed scientific publications, containing multilingual data used for machine translation. Available at https://huggingface.co/datasets/qanastek/WMT-16-PubMed. License CC-BY-4.0.

HAL: Corpus extracted from the HAL platform, grouping French scientific publications in the biomedical domain. NACHOS versioning. Available via harvesting following the api protocol https://api.documentation-administrative.gouv.fr/oai. License Etalab.

HAS: Data from the High Authority of Health. NACHOS versioning. Available at https://www.data.gouv.fr/fr/datasets/textes-des-publications-de-la-has-7/. License Etalab.

QUAERO: Corpus of multilingual medical documents from MEDLINE titles and documents from the European Medicines Agency (EMEA-V3), used for training and evaluating models of automatic medical language processing. NACHOS versioning. Available at https://huggingface.co/datasets/DrBenchmark/QUAERO. License GNU Free Documentation License.

WMT18_MEDLINE: Corpus of biomedical texts from Medline, used in the context of the WMT18 challenge for automatic translation. NACHOS versioning. Available at https://www.statmt.org/wmt18/biomedical-translation-task.html. License CC BY-NC-SA 3.0, CC BY-NC-ND 4.0.

ISTEX: Corpus of scientific publications from the ISTEX platform, gathering French scientific literature. NACHOS versioning. Available at https://data.istex.fr/. License Etalab.

CLEAR: Corpus containing texts from 3 sources: encyclopedia, pharmaceutical notices and medical article abstracts. NACHOS versioning. Available at https://shs.hal.science/halshs-01968355. Research-only license.

MANTRA_GSC: Dataset extracted from biomedical corpora (Medline abstract titles, pharmaceutical notices, biomedical patents), with independent concept annotation according to a subset of the UMLS. NACHOS versioning. Available at https://huggingface.co/datasets/bigbio/mantra_gsc. License CC-BY-4.0.

Preprocessing steps

Text cleaning

All the documents were preprocessed using a pipeline inspired by FlauBERT (Le et al., 2020), including Unicode conversion and normalization, removal of characters outside standard French encoding, removal of multiple spaces, and removal of URLs.

To this initial cleaning script, additional steps were added due to the lack of relevant content in some documents included in the corpus. These were based on criteria such as a minimum word count (=5; a higher number would have been too restrictive for dialogues) in the texts that were retained.

De-duplication

To avoid overfitting on redundant samples in our dataset, we added an additional deduplication step during preprocessing. We used a very “classic” method based on MinHash similarity, with a similarity threshold of 0.85 and a number of permitted permutations set to 128.

This deduplication was applied during the transfer of the sourced datasets to the ready-to-use, unsourced corpus. Indeed, since some corpora intersect, the granularity of the source becomes less relevant because the documents are compared in an inter-corpus manner.

Features Scheme

Column Name	Data Type	Description
instruction	string	instruction-tuning only feature, corresponding to the system prompt for instruction-tuning samples.
input	string	input text, regardless of the adaptation method (e.g., finetuning or instruction-tuning). For instruction-tuning, this is the "user prompt" or "question".
output	string	instruction-tuning only feature gold standard output for supervised instruction-tuning.
source	string	dataset name of the data sample.
document_type	string	typology of document (e.g., Scientific, Encyclopedic, Clinical, Medication, Question-Answering, Dialogue, Regulation).

Statistics

Document-type granularity

FINETUNING data

	nb_docs	nb_words	mean_words	std_words	nb_chars	mean_chars	std_chars
Total	905342	9.00141e+08	994.255	6719.46	5.61243e+09	6199.24	41099.6
Scientific	640313	8.49585e+08	1326.83	7932.88	5.27754e+09	8242.13	48478.3
Medicinal	233960	2.44849e+07	104.654	647.2	1.63167e+08	697.415	4332.35
Clinical	16100	1.75665e+07	1091.08	1290.35	1.15255e+08	7158.72	8430.4
Encyclopedic	9957	6.53102e+06	655.923	1252.04	4.32721e+07	4345.89	8209.94
Education	22	1.71519e+06	77963.1	47413.5	1.16235e+07	528341	321525
Question Answering	275	111792	406.516	264.436	626549	2278.36	1402.57
Regulation	1111	70081	63.0792	54.7356	478447	430.645	365.089
Medical	2152	42460	19.7305	13.3516	280626	130.402	92.0109
Dialogue	1452	34044	23.4463	73.5192	188202	129.616	394.801

INSTRUCTION-TUNING data

	nb_docs	nb_words	mean_words	std_words	nb_chars	mean_chars	std_chars
Question Answering	22390	1.78385e+06	79.6716	59.3966	1.17989e+07	526.971	372.088
Total	22390	1.78385e+06	79.6716	59.3966	1.17989e+07	526.971	372.088

Source-wise granularity

FINETUNING data

	nb_docs	nb_words	mean_words	std_words	nb_chars	mean_chars	std_chars
Total	905342	9.00141e+08	994.255	6719.46	5.61243e+09	6199.24	41099.6
HAL	26987	7.03474e+08	26067.1	26603.8	4.32567e+09	160287	160053
HAS	11334	9.61734e+07	8485.39	16098.9	6.20009e+08	54703.4	102858
ISTEX	12179	4.31384e+07	3542.03	2156.57	2.82624e+08	23205.9	14238.5
BDPM	11023	2.00358e+07	1817.63	2409.58	1.35081e+08	12254.5	16062.4
E3C	7499	1.58646e+07	2115.57	1222.36	1.0414e+08	13887.2	7923.95
WIKIPEDIA	9957	6.53102e+06	655.923	1252.04	4.32721e+07	4345.89	8209.94
WMT16	587563	6.49552e+06	11.055	5.40785	4.73973e+07	80.6676	37.5056
EMEA_V3	222937	4.44909e+06	19.9567	15.5252	2.80864e+07	125.984	99.953
CERIMES	22	1.71519e+06	77963.1	47413.5	1.16235e+07	528341	321525
FRASIMED	2048	1.3229e+06	645.945	333.9	8.73338e+06	4264.34	2207.72
CAS	712	232389	326.389	242.842	1.52772e+06	2145.68	1501.74
CLEAR	6	226123	37687.2	46388.3	1.36912e+06	228188	280743
ESSAI	5841	146530	25.0865	14.2491	854518	146.297	83.1409
DEFT2021	275	111792	406.516	264.436	626549	2278.36	1402.57
QUAERO	2083	66877	32.1061	161.208	394933	189.598	905.512
CNEDIMTS	813	58345	71.7651	60.599	398478	490.133	403.23
ECDC_TM	2152	42460	19.7305	13.3516	280626	130.402	92.0109
PXCORPUS	1414	18372	12.9929	6.0802	103531	73.2185	33.7791
MQC	38	15672	412.421	223.131	84671	2228.18	1179.41
QUALISCOPE	298	11736	39.3826	19.5879	79969	268.352	131.707
WMT18_MEDLINE	49	7719	157.531	65.3727	51627	1053.61	416.966
MANTRA_GSC	112	3085	27.5446	39.6518	22356	199.607	306.097

INSTRUCTION-TUNING data

	nb_docs	nb_words	mean_words	std_words	nb_chars	mean_chars	std_chars
Total	22390	1.78385e+06	79.6716	59.3966	1.17989e+07	526.971	372.088
MEDIQAL	19907	1.6593e+06	83.3526	61.6255	1.09334e+07	549.225	386.325
FRENCHMEDMCQA	2483	124547	50.1599	19.6412	865475	348.56	126.799

File Organization

PARCOMED_research_only/
├── fine-tuning/
│   ├── dataset1_part1.parquet
│   ├── dataset1_part2.parquet
│   └── ...
├── instruction-tuning/
│   ├── dataset2_part1.parquet
│   ├── dataset2_part2.parquet
│   └── ...
└── README.md

Usage

from dataset import load_dataset

data = load_dataset(
    "HealthDataHub/PARCOMED_research_only", 
    split="train", 
    data_dir="finetuning"  # or "instruction-tuning"
    download_mode="force_redownload", 
    verification_mode="no_checks", 
)

Contribution

This dataset was created thanks to the the collaborative effort of PARTAGES development teams, including data identification, collection and licensing analysis. We acknowledge and thank all individuals and teams involved in its creation, more specifically:

• Armand VIOLLE, Stéphane OHAYON, Chaïma ABDELLAOUI and Xavier TANNIER from LIMICS (Sorbonne Université)
• Aidan MANNION, Cécile MACAIRE, Didier SCHWAB, Lorraine GOEURIOT and François PORTET from LIG (Université Grenoble Alpes, CNRS, Grenoble INP)
• Pierre ZWEIGENBAUM, Aurélie NÉVÉOL, Christophe SERVAN from LISN (CNRS, Université Paris-Saclay, INRIA)
• Ygor GALLINA, Mickaël ROUVIER from LIA (Avignon Université)