Dataset Card for PARHAF-pseudo-annotated

Reporting Issues & Contributing

If you encounter any errors or inconsistencies in this dataset, please report them in the discussion section of the "Community" tab on Hugging Face.

For more substantial contributions or collaboration opportunities, feel free to contact us directly.

Dataset Summary

PARHAF-pseudo-annotated is a subpart of the PARHAF corpus, 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 for the pseudonymization of identifying entities in clinical reports..

This dataset contains training data only. The test set will remain under embargo to enable future evaluations under controlled conditions, limiting the risk of LLM contamination through prior data exposure. Please contact us for access to the test data.

This training dataset is divided into a train split (80%) and a dev split (20%) to facilitate experimental design and reproducibility across teams. Teams are free to use the full training set or define a different split configuration.

Each patient record was:

• written by a senior medical resident
• reviewed by another senior medical resident, from the same specialty
• annotated by a specialist of the use case
• curated by another specialist of the use case

Data statistics

DATASET SUMMARY

Indicator	Value
Complete Dataset
Number of JSON files	509
Total annotations	6976
Average document length	4021 characters
80% Threshold (annotations)	5580

Complete dataset distribution by type

Type	Annotations	% of total
EntiteAnonymisation	6976	100.00%

Complete dataset distribution by populated field

Type.Field	Occurrences	% of total annot.
EntiteAnonymisation.Category	6976	100.00%
EntiteAnonymisation.RoleLOC	87	1.25%
EntiteAnonymisation.RoleNUM	8	0.11%
EntiteAnonymisation.RolePER	2466	35.35%

Train / Dev Split Results

Indicator	TRAIN	DEV
Number of files	408	101
Number of annotations	5580	1396
Percentage of dataset	79.99%	20.01%
Avg. document length (chars)	3957	4280

Train / Dev distribution by type

Type	TRAIN	DEV	% train of type
EntiteAnonymisation	5580	1396	79.99%

Train / Dev distribution by populated field

Type.Field	TRAIN	DEV	% train of field
EntiteAnonymisation.Category	5580	1396	79.99%
EntiteAnonymisation.RoleLOC	66	21	75.86%
EntiteAnonymisation.RoleNUM	5	3	62.50%
EntiteAnonymisation.RolePER	1971	495	79.93%

Train / Dev distribution by value

Type.Field.Value	TRAIN	DEV	% train
EntiteAnonymisation.Category.IDENTIFYING_DATE	2312	568	80.28%
EntiteAnonymisation.Category.LAST_NAME	1111	287	79.47%
EntiteAnonymisation.RolePER.Patient	978	257	79.19%
EntiteAnonymisation.RolePER.Carer	973	232	80.75%
EntiteAnonymisation.Category.FIRST_NAME	860	208	80.52%
EntiteAnonymisation.Category.FAMILY_STATUS	767	203	79.07%
EntiteAnonymisation.Category.PATIENT_SOCIAL_IDENTITY	424	99	81.07%
EntiteAnonymisation.Category.CITY	43	11	79.63%
EntiteAnonymisation.RoleLOC.Patient	33	13	71.74%
EntiteAnonymisation.Category.COUNTRY	19	9	67.86%
EntiteAnonymisation.RolePER.Other	20	6	76.92%
EntiteAnonymisation.RoleLOC.Hospital	18	4	81.82%
EntiteAnonymisation.Category.UNIDENTIFYING_DATE	19	0	100.00%
EntiteAnonymisation.RoleLOC.Other	15	4	78.95%
EntiteAnonymisation.Category.PATIENT_BIRTHDATE	12	5	70.59%
EntiteAnonymisation.Category.PHONE_NUMBER	5	3	62.50%
EntiteAnonymisation.Category.PATIENT_NATIONALITY	4	2	66.67%
EntiteAnonymisation.RoleNUM.Carer	4	1	80.00%
EntiteAnonymisation.Category.ADDRESS	3	0	100.00%
EntiteAnonymisation.RoleNUM.Hospital	1	1	50.00%
EntiteAnonymisation.Category.URL	1	1	50.00%
EntiteAnonymisation.RoleNUM.Other	0	1	0.00%

Data Origin

The clinical reports are extracted from the PARHAF corpus. Please refer to PARHAF documentation for more information about this corpus.

Languages

• fr_FR

Dataset Structure

We distribute both a Hugging Face dataset and a standalone version of the corpus. The standalone dataset consists of a JSON file per patient report, in UIMA CAS JSON format. This format constitutes 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.

One dataset instance corresponds to one report.

Hugging Face dataset

This snippet shows how to extract and iterate over medical report information per patient using the datasets library.

 import pandas as pd
 from datasets import load_dataset
 

dfs = {cfg: load_dataset("HealthDataHub/PARHAF-pseudo-annotated", cfg, split="train").to_pandas()
       for cfg in ["document_metadata", "spans"]}

for patient_raw in dfs["document_metadata"].itertuples():
    report_id = patient_raw.report
    text = patient_raw.full_text
    report_spans = dfs["spans"][dfs["spans"]["report"] == report_id]
    ...  

Data Fields

Path	Type	Description	Possible values
document_metadata
report	string	Identifiant unique du rapport
full_text	string	Texte intégral du rapport
spans
report	string	Identifiant du rapport
span_id	integer	Identifiant de l'annotation
span_type	string	Type de l'entité annotée	EntiteAnonymisation
begin	integer	Offset de début
end	integer	Offset de fin
span_text	string	Texte de l'entité
attribute_Categorie	string	Categorie	ADDRESS, CITY, COUNTRY, FAMILY_STATUS, FIRST_NAME, IDENTIFYING_DATE, LAST_NAME, PATIENT_BIRTHDATE, PATIENT_NATIONALITY, PATIENT_SOCIAL_IDENTITY
attribute_RolePER	string	RolePER	Carer, Other, Patient
attribute_RoleLOC	string	RoleLOC	Hospital, Other, Patient
attribute_RoleNUM	string	RoleNUM	Carer, Hospital

Data Splits

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.

Annotation Guidelines

You can find the detailed annotation protocol here: annotation_guidelines.pdf

Licensing Information

This dataset is released under licenses:

• CC BY 4.0
• Etalab 2.0

Citation Information

[More Information Needed]