README.md

metadata

features:
  - name: id
    dtype: string
  - name: local_id
    dtype: string
  - name: specialty
    dtype: string
  - name: author
    dtype: string
  - name: reviewer
    dtype: string
  - name: pool
    dtype: string
  - name: suggested_scenario
    struct:
      - name: name
        dtype: string
      - name: age
        struct:
          - name: value
            dtype: int32
          - name: unit
            dtype: string
      - name: sex
        dtype: string
      - name: admission_mode
        dtype: string
      - name: discharge_mode
        dtype: string
      - name: primary_procedure
        struct:
          - name: code
            dtype: string
          - name: description
            dtype: string
      - name: primary_diagnosis
        sequence:
          - name: code
            dtype: string
          - name: description
            dtype: string
      - name: type_of_care
        dtype: string
  - name: documents
    sequence:
      - name: type
        dtype: string
      - name: header
        dtype: string
      - name: text
        dtype: string
      - name: word_count
        dtype: int32
  - name: structured_abstract
    struct:
      - name: primary_diagnosis
        sequence:
          - name: code
            dtype: string
          - name: description
            dtype: string
      - name: primary_procedure
        sequence:
          - name: code
            dtype: string
          - name: description
            dtype: string
      - name: admission_mode
        dtype: string
      - name: discharge_mode
        dtype: string
      - name: length_of_stay
        struct:
          - name: value
            dtype: int32
          - name: unit
            dtype: string
config_name: default
splits:
  - name: train
    num_bytes: 21325362
    num_examples: 4254
download_size: 0
dataset_size: 21325362

Dataset Card for PARHAF

Table of Contents

• Dataset Description
  • Dataset Summary
  • Languages
• Dataset Structure
  • Standalone corpus (out of Hugging Face)
  • Hugging Face dataset
  • Data Instances
  • Data Fields
  • Controlled Vocabularies
  • Data Splits
• Dataset Creation
• Changelog
  • 1.0
• Additional Information
  • Licensing Information
  • Citation Information

Dataset Description

Note for users interested in the PARTAGES use cases on pseudonymisation, coding, oncology, and infectiology: This dataset contains unlabeled data only. Labeled versions are available in separate datasets on the same platform. Besides, the documents for the test set of all use cases are excluded. They will remain under embargo to enable future evaluations under controlled conditions, limiting the risk of LLM contamination through prior data exposure. Please contact us if you wish to get access to the test set to evaluate your system.

Paper: https://arxiv.org/pdf/2603.20494

Dataset Summary

PARHAF is an open French corpus of human-authored clinical reports of fictional patients. It was created to support the development and evaluation of clinical NLP systems under strict health-data protection constraints. Each patient is each documented with structured clinical information (diagnosis, procedures, care pathway, discharge data).

Each patient record was written by a senior medical resident and reviewed by another senior medical resident, from the same specialty. You can consult the guidelines provided to the writers and reviewers (in French).

Data statistics

Main statistics

Patients	Documents	Words
4259	6190	3952583

Patient count per specialty

Specialty	General	CU 2 - ICD-10 coding	CU 5a - Oncology (biomarkers)	CU 5b - Oncology (response to treatment)	CU 6 - Infectiology	Total
Cardiology	96	0	0	0	0	96
Cardiovasc. Surg.	39	0	0	0	0	39
Critical Care	44	0	0	0	0	44
Digestive Surg.	170	125	0	0	0	295
Gastro-Hepatology	170	0	0	0	0	170
Gen. Internal Med.	82	0	0	0	0	82
Geriatrics	155	0	0	0	0	155
Gynecology	212	0	0	0	0	212
Hematology	153	0	0	0	0	153
Infectious Dis.	151	0	0	0	134	285
Internal Med.	228	0	0	0	0	228
Nephrology	104	0	0	0	0	104
Neurology	251	0	0	0	0	251
Obstetrics	174	0	0	0	0	174
Oncology	399	0	0	108	0	507
Ortho & Trauma Surg.	143	98	0	0	0	241
Pathology	285	0	152	0	0	437
Pediatrics	82	0	0	0	0	82
Pulmonology	531	0	0	0	0	531
Urology	108	65	0	0	0	173
Total	3577	288	152	108	134	4259

Note: In case you wonder why CU 1 (Pseudonymisation) is not listed below: The patients for this use case have been sampled from the general distribution, that's why they don't appear here. To know which documents are part of this use case, consult the specific dataset PARCOMED.

Data Origin

The clinical reports were authored by senior medical residents specifically for this corpus.

The source data for building clinical scenarios comes from National Hospital claims data available in the SNDS national French database. To comply with privacy regulations regarding the use of SNDS data, scenarios are built by sampling on the observed distributions. The diagnosis distribution aims at reducing the over-representation of very common conditions and including less frequent situations.

The writers were provided with a scenario made of:

• A primary diagnosis
• A patient age
• An admission mode (when relevant)
• A discharge mode (when relevant)
• A type of care (when relevant)

A list of common secondary diagnoses for the specified primary diagnoses was also provided for information.

A standardized clinical report template was also provided to the writers. It defined a generic structure designed to reflect the expected logic of clinical documentation. It includes the core sections of a medical report (such as patient history, clinical findings, assessment, and care plan) while remaining flexible through adaptable templates based on the report type, medical specialty, and identified use cases.

Intended usages

This dataset can be used to support a variety of applications, including:

• Sharing clinical notes and annotations
• Pooling efforts within the clinical NLP community
• Benchmarking French medical LLMs
• Enabling reproducible clinical NLP research
• Supporting medical teaching and education
• Promoting work on PARTAGES’s 7 use cases
• Enabling privacy-safe data augmentation

Excluded usages

The dataset is not intended to be used for

• Clinical decision-making or patient care (no diagnostic, prognostic, or therapeutic use)
• Clinical validation or performance claims (models trained on this data are not clinically validated)
• Generalization to unseen hospitals, regions, or practices
• Epidemiological or population-level inference (≠ real-world prevalence or distributions)
• Assessing real-world safety or clinical risk (rare adverse events, edge cases)
• Replacing real clinical data for deployment (research-only)
• Stress-testing models on realistic clinical language (no time pressure, workload, interruptions)

This dataset is limited to hospital-based clinical documentation and focuses on selected document types that are central to the patient record but not representative of the full range of medical documentation. It excludes several important categories, such as imaging reports, prescriptions, referral letters, and procedural or operative reports, and therefore does not capture the full diversity of clinical documentation.

Languages

• fr_FR

Dataset Structure

We distribute both a Hugging Face dataset and a standalone version of the corpus. The standalone JSON files constitute the canonical version of the corpus. The Hugging Face dataset (Parquet/Arrow) is a derived representation generated automatically from the JSON files.

Both formats therefore contain identical information and differ only in storage layout.

Standalone corpus (out of Hugging Face)

The standalone dataset consists of a single JSON file containing structured metadata about fictitious patients and the clinical documents associated with them. Each entry in the data array corresponds to one patient record and includes patient-level metadata, contextual information about the care scenario, and a list of associated documents.

The documents themselves are not embedded in the JSON file. Instead, each document is referenced via a relative file path pointing to an external text file. These text files are stored separately and organized by medical specialty, with one directory per specialty. Each document file contains raw, unannotated plain text in French, with no markup, labels, or structural tags applied.

The JSON file, therefore, serves as the index and metadata layer for the corpus, while the directory structure contains the raw textual content. The linkage between metadata and text relies exclusively on the relative paths specified in the documents[].path fields.

import json
import os
with open(json_path, "r") as f:
    data = json.load(f)
    
patients = data["data"]
patient = patients[0]       # First patient
patient_id = patient["id"]
specialty = patient["specialty"]
scenario = patient["suggested_scenario"]
patient_fake_name = scenario["name"]
primary_diagnosis = scenario["primary_diagnosis"]
patient_reports = patient["documents"]
...

# Each patient can have one, two or three reports
for report in patient_reports:
    type = report["type"]
    path = report["path"]
    # The content of the report is not in the JSON file
    # but in distinct text files
    abs_path = os.path.join(os.path.dirname(json_path), path)
    with open(abs_path, "r") as f:
        text = f.read()
    print(type, text)
    ...

Hugging Face dataset

The Hugging Face dataset and the standalone JSON file contain the same information but use different data representations. First, in the Hugging Face dataset (parquet format), the textual content of the document is embedded in the dictionary. Second, the JSON standalone representation follows a conventional nested structure (a list of objects, each containing a list of {type, text} dictionaries), as described below in the description of data instances.

In contrast, the Hugging Face version is stored using a columnar format optimized for efficient loading and machine learning. As a consequence, sequences of dictionaries are represented as a dictionary of lists (one list per field) rather than a list of dictionaries. This structural change does not alter the data content, only its storage layout.

from datasets import load_dataset

ds = load_dataset("HealthDataHub/PARHAF")
patient = ds["train"][0]       # First patient
patient_id = patient["id"]
specialty = patient["specialty"]
scenario = patient["suggested_scenario"]
patient_fake_name = scenario["name"]
primary_diagnosis = scenario["primary_diagnosis"]
patient_reports = patient["documents"]
...
# In Hugging Face dataset, sequences of dictionaries are represented 
# as a dictionary of lists (one list per field) rather than a list of dictionaries.
types = patient_reports["type"]
texts = patient_reports["text"]
for type, text in zip(types, texts):
    print(type, text)
    ...

Data Instances

One dataset instance corresponds to one patient and contains all associated documents and metadata.

Each example includes:

• id: patient identifier
• local_id
• specialty
• author
• reviewer
• pool
• suggested_scenario: structured clinical metadata
• documents: list of clinical reports with full text
• structured_abstract: optional structured summary

The dataset preserves a hierarchical patient-level organization. Note that this corresponds to the standalone dataset; As mentioned above, the HF dataset is stored using a columnar format.

patient
 ├─ id: string
 ├─ local_id: string
 ├─ specialty: string
 ├─ author: string
 ├─ reviewer: string
 ├─ pool: string
 ├─ suggested_scenario:
 │   ├─ name: string
 │   ├─ age:
 │   │   ├─ value: integer
 │   │   └─ unit: string
 │   ├─ sex: string
 │   ├─ admission_mode: string (optional)
 │   ├─ discharge_mode: string (optional)
 │   ├─ primary_procedure: {code, description} (optional)
 │   ├─ primary_diagnosis: {code, description}
 │   └─ type_of_care: string (optional)
 ├─ documents: [document]
 │   └─ document
 │       ├─ type: string
 │       ├─ header: string
 │       ├─ path: string
 │       └─ word_count: integer
 └─ structured_abstract (optional)
     ├─ primary_diagnosis: [{code?, description}]
     ├─ primary_procedure: [{code, description}]
     ├─ admission_mode: string
     ├─ discharge_mode: string
     └─ length_of_stay: {value, unit}

Data Fields

Field names use dot notation to describe nested objects.

data[](one patient)

Path	Type	Description
id	string	Global unique patient identifier
local_id	string	Local identifier within the specialty
specialty	string	Medical specialty
author	string	Author trigram
reviewer	string	Reviewer trigram
pool	string	Dataset partition. Some patients belong to specific use cases (CU1–CU6)
suggested_scenario	dict	Scenario provided to the report writer
documents[]	array	List of reports for this patient
structured_abstract	dict	Optional uncurated abstract written by the author. May not match report content

suggested_scenario

Path	Type	Description
name	string	Fictional patient name
age.value	integer	Age value
age.unit	string	Age unit
sex	string	Patient sex
admission_mode	string	Admission source
discharge_mode	string	Discharge destination
primary_procedure.code	string	CCAM code
primary_procedure.description	string	Procedure label
primary_diagnosis.code	string	ICD-10 code
primary_diagnosis.description	string	Diagnosis label
type_of_care	string	Care description

documents[](patients reports, 1-3 per patient)

Path	Type	Description
type	string	Document type
header	string	Document title
word_count	integer	Number of words in the report
path	string	Relative path to raw text

Controlled Vocabularies

The following fields use closed sets of values and should be treated as categorical variables rather than free text.

specialty

ANATOMOPATHOLOGIE
CANCERO ADULTE
CARDIOLOGIE
CHIR ORTHO ET TRAUMATO
CHIR.CARDIO-VASC.
CHIRURGIE VISCERALE
GYNECOLOGIE
HEMATOLOGIE CLINIQUE
HEPATO-GASTRO-ENTERO
MALADIES INFECTIEUSES
MEDECINE GERIATRIQUE
MEDECINE INTER-SPECIALITES
MEDECINE INTERNE
MEDECINE PEDIATRIQUE
NEPHROLOGIE
NEUROLOGIE
OBSTETRIQUE
PNEUMOLOGIE
REANIMATION
UROLOGIE

pool

'CU 1 - Pseudonymisation'
'CU 2 - ICD-10 coding' 
'CU 5a - Oncology (biomarkers)', 
'CU 5b - Oncology (response to treatment)', 
'CU 6 - Infectiology', 
'General'

age.unit

'ans', 
'mois'

sex

'F',
'M'

admission_mode

'admission non programmée suite à un contact avec le médecin traitant dans les 48h',
'admission non programmée suite à un contact avec un médecin suivant le patient en consultation (médecin traitant par exemple) dans les 48h',
'admission par les urgences',
'domicile',
'domicile (à modifier pour entrée par les urgences)',
'domicile sans passer par le service des urgences',
'domicile- service de gastro-entérologie',
'entrée par les urgences',
"entrée par un autre service d'hospitalisation",
"transfert d'un autre hôpital",
'transfert depuis un autre service',
'transfert par le service des urgences',
'urgences'
None

discharge_mode

'admission en smr',
'domicile',
'décès',
'retour au soins de suite et réadaptation',
'retour à domicile',
'retour à domicile -> hématologie stérile',
'retour à domicile ; transfert gastrologie',
'retour à domicile > changer pour transfert en chirurgie digestive',
'retour à domicile → hospitalisation',
'sortie en soins de suite',
'ssr neurologique',
'transfert',
'transfert dans un autre hôpital',
'transfert dans un autre service',
'transfert en gastro-entérologie',
'transfert en gynécologie',
'transfert en psychiatrie',
'transfert en soins de suite et réadaptation',
'transfert en soins de suite et rééducation'
None

Document's type

'ACCOUCHEMENT',  # Birth 
'ANAPATH',       # Pathology report
'CRC',           # CR consultation
'CRH',           # CR hospitalisation
'CRO',           # CR opération
'MATERNITE',     # Maternity
'URGENCES'       # ER

Data Splits

Note for users interested in the PARTAGES use cases on pseudonymisation, coding, oncology, and infectiology: This dataset contains unlabeled data only. Labeled versions are available in separate datasets on the same platform. Besides, the documents for the test set of all use cases are excluded. They will remain under embargo to enable future evaluations under controlled conditions, limiting the risk of LLM contamination through prior data exposure. Please contact us if you wish to get access to the test set to evaluate your system.

Dataset Creation

PARTAGES releases an open corpus of human-authored medical reports in French that will enable players from industry and academia to address specific use cases based on general medical generative LLMs.

This was made possible by:

• A partnership with over 17 associations and unions of medical residents, in multiple medical and surgical specialties : 120 medical residents were recruited as writers;
• The creation of guidelines to choose clinical cases based on Data from the National Health Data System (SNDS) as support scenarios to author medical reports;
• The authoring by medical residents of synthetic medical reports;
• The creation of an open-source corpus in French to train and evaluate language models on specific use cases.

Changelog

1.0

The currently delivered version includes all documents that are not assigned to a specific clinical use case. Documents related to specific use cases will be delivered at a later stage.

Only the training set is released here. The remaining portion of the corpus will be temporarily embargoed to enable future evaluations under controlled conditions, thereby limiting the risk of large language model contamination through prior exposure to the data. You can evaluate your system on the test set through the CodaBench platform.

Please contact us for information about how to evaluate your system on our unreleased test data

Additional Information

Licensing Information

This dataset is released under licenses:

• CC BY 4.0
• Etalab 2.0

Citation Information

[More Information Needed]