Datasets, benchmarks and open models for studying the French press.
Non-profit association under French law (loi 1901) · Paris · RNA W751279600
Supported by
Le French News Lab explores how French news shapes political life. Our work looks at the stories, frames, and language that structure public debate, combining computational methods with media and political analysis.
We develop datasets, methods, and public-facing research that make the French media ecosystem easier to study, question, and understand.
Grants, papers and data releases from the lab, newest first.
All newsAnnouncement
GENCI has awarded Le French News Lab computing time on Jean Zay, bringing France’s national supercomputing infrastructure to our research on French news.
Read the announcementFrIdéo and Framing by Wording, Framing by Selection, both accepted for oral presentation.
87,637 articles from 13 publishers, harmonised into 7 shared editorial desk categories.
Scales, corpora and framing audits for the French press — each one documented enough that a disagreement about the media can become a disagreement about data.
Ideology scale
FrIdéo places French national news outlets on a single ideology scale by fusing nine independent evidence families under empirical Bayes shrinkage. Each placement decomposes into the evidence that produced it; leave-one-out ablation holds at ρ ≥ 0.992.
Accepted for oral presentation at the 9th International Conference on Natural Language and Speech Processing, Trento, September 2026.
Headline framing
To our knowledge the largest framing audit of the French press: wording and selection measured separately across 902,111 headlines from 25 outlets, 2022–2025. A released 10,000-headline supervision set carries per-item annotator agreement rather than a collapsed consensus.
Accepted for oral presentation at the 9th International Conference on Natural Language and Speech Processing, Trento, September 2026.
Role framing
In 28,592 headlines from 25 outlets, La France insoumise is cast as a conflict actor and Rassemblement National as a strategic competitor. The asymmetry holds after outlet and year controls; at Le Journal du Dimanche it inverted after the 2023 editorial takeover.
Editorial desk classification
87,637 articles from 13 publishers, mapped onto seven editorial desks. Where a URL slug is unambiguous the label is deterministic (72.2%); the ambiguous remainder is LLM-labelled and audited against a blinded human sample at κ = 0.806. The CamemBERT reference model reaches macro-F1 0.847 on the held-out split.
We build the measurement infrastructure that makes the French news ecosystem possible to study: annotated corpora, reproducible benchmarks, open models, and peer-reviewed methods for questions that are usually argued about without evidence.
Claims about bias, framing and slant in the French press are usually made without a number anyone can check. We build scales and benchmarks that are documented, versioned and auditable, so that a disagreement about the media can become a disagreement about data.
Papers are published with the datasets behind them. Results are reported as means across multiple random seeds rather than a single best run, because a number that only appears once is not a result.
Our datasets and models are published on Hugging Face, data.gouv.fr and Kaggle under open licences, free to use. Nothing on this site is behind a paywall or a signup form.
Everything we build for a paper is released alongside it. These are the current public datasets and models — free to download, with data dictionaries and full metadata.
A cross-publisher corpus of French news articles annotated by editorial desk, built so that topic classifiers can be tested on publishers they were never trained on — the case where reported accuracy usually falls apart. Released with the train and test splits used in the paper.
From the paper →A CamemBERT-base classifier fine-tuned on FrenchNews-7 for editorial desk classification, published with its model card and the per-seed evaluation scores reported in the paper.
Continuous ideology scores for thirty French national news outlets, fusing nine independent families of evidence rather than resting on any single one. Published as CSV and JSON with a data dictionary and full metadata, under CC BY 4.0.
From the paper →We collaborate with researchers, journalists, civil-society organisations and academic institutions working on the French information ecosystem. If you want to use our data, replicate a result, or propose a study, get in touch — we answer every message.
