finemed-rephrased-fr

What finemed-rephrased-fr is

FineMed-rephrased-fr is a large-scale corpus of synthetically rewritten French medical text created by doctolib-lab. It consists of 13.6 million documents comprising 4.53 billion words derived from FineMed-fr, an unfiltered medical dataset extracted from FineWeb-2, FinePDFs, and FineWiki. The dataset was designed to serve alongside the original source texts for pretraining French biomedical encoders such as DoctoBERT and DoctoModernBERT.

Web-scraped biomedical corpora often contain extensive non-medical filler, uninformative boilerplate, and optical character recognition errors that standard filtering leaves intact. FineMed-rephrased-fr addresses this by prompting an open-weights large language model to faithfully rephrase raw sources across diverse stylistic registers, audiences, and genres. This process eliminates non-medical filler while retaining factual clinical entities, more than doubling overall medical-term density compared to the original corpus.

The dataset provides both the newly generated texts and the source inputs, complete with recomputed quality ratings and medical entity annotations. This enables researchers to examine how synthetically altered term co-occurrences, register variations, and entity-dense reformulations influence downstream clinical language modelling.

How it was made

The rephrasing pipeline follows a two-stage medical adaptation of the Massive Genre-Audience (MGA) framework powered by Qwen3.5-35B-A3B. Prior to generating text, a coarse pre-screen filters FineMed-fr annotations to avoid processing texts that lack substantial medical subject matter. Stage 1 executes a medical-content gate to discard unsuitable documents, independently brainstorms candidate genres and audiences, pairs them for diversity, and samples one pair per document. Stage 2 rewrites the source for the assigned pair, eliminating non-medical filler while strictly preserving clinical entities, numbers, and facts. To introduce surface variation and mirror authentic clinical communication, the generation specifies formal versus telegraphic registers and controls abbreviation density across expanded, moderate, or heavy settings. Non-medical personally identifiable information is replaced with fictionalised values.

After generation, texts undergo quality filtering via DataTrove to enforce French language identification and remove repetitive text with a Gopher repetition filter. The final records are re-annotated for educational quality and medical-term density across eight Unified Medical Language System (UMLS)-adapted classes, while the underlying 15-class medical subdomain labels are carried over directly from the source.

Key facts

What you can use it for

Biomedical Language Model Pretraining

Use the corpus alongside raw web text to train French clinical encoders such as DoctoBERT, exposing the models to dense medical vocabulary and diverse syntactic structures.

Medical Register and Style Adaptation

Train text-generation models to translate patient-facing communications, narrative forums, or journal literature into telegraphic clinical notes and standardised drug monographs.

Clinical Named Entity Recognition (NER)

Benchmark and train French entity extractors using parallel spans mapped across eight entity classes between original and rephrased documents.

Data De-identification Robustness

Evaluate de-identification pipelines on synthetically injected fictional names and addresses juxtaposed with authentic medical details.

Explore the data

The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.

Strengths and limits

Questions