--- pretty_name: FrenchNews-7 language: - fr license: cc-by-4.0 task_categories: - text-classification task_ids: - multi-class-classification size_categories: - 10K **Note on the dataset viewer.** The manifest ships as a single CSV; the `split` column carries > the partition. The Hub viewer therefore shows all 87,637 rows under one `train` view. Filter on > the `split` column to recover the three partitions: > > ```python > from datasets import load_dataset > ds = load_dataset("LeFrenchNewsLab/frenchnews-7", split="train") > test = ds.filter(lambda r: r["split"] == "test") > ``` ## Label Distribution | Label | Articles | |---|---:| | Société | 19,324 | | International | 18,642 | | Culture & Loisirs | 18,350 | | Politique | 10,546 | | Sport | 8,404 | | Économie | 8,174 | | Sciences & Technologies | 4,197 | ## Annotation Labels are assigned by a two-bucket pipeline that prefers deterministic assignment over model judgment wherever possible. | Bucket | Method | Articles | Share | |---|---|---:|---:| | A | Deterministic slug rules (74 rules) | 63,302 | 72.2% | | B | LLM zero-shot (temperature = 0) | 24,335 | 27.8% | **Bucket A** matches the **first URL path segment** against a 74-rule lookup table (`annotation/bucket_a_slug_rules.json`). Matching is **case-insensitive** — several publishers emit capitalised segments such as `/Culture/`, `/Societe/` and `/International/`. Reproducing with case-sensitive matching yields 58,926 instead of 63,302. **Bucket B** covers articles whose slug carries no unambiguous desk signal (`/idees`, `/debats`, `/story`, `/flash-actu`, regional or absent slugs). ### Annotation quality A four-rater agreement study was run on a 300-article stratified random sample of Bucket B (GPT-OSS-120B as original labeller, Gemini-2.5-flash-lite as independent second LLM, an adjudicator, and a blinded second human annotator shown headline and body only — no URLs, outlet names, or model labels). | Rater pair | Agreement | Cohen's κ | |---|---:|---:| | Blinded human ↔ reference labels | 84.0% | 0.806 | | GPT-OSS ↔ blinded human | 80.7% | 0.766 | | Original LLM ↔ Gemini | 87.3% | 0.846 | | Original LLM ↔ adjudicated gold | 95.3% | 0.943 | | Gemini ↔ adjudicated gold | 91.3% | 0.895 | Fleiss' κ = 0.835 across all four raters; three-rater Krippendorff's α = 0.789. Bucket B labels are functionally inert for reported results: training on Bucket A only shifts held-out macro-F1 by −0.004 (95% CI [−0.024, +0.024]; McNemar p = 0.63). ## Publishers - 20 Minutes - JDD - L'Express - L'Humanité - La Croix - Le Figaro - Le HuffPost - Le Monde - Le Parisien - Le Point - Ouest-France - Slate.fr - TF1 INFO ## Manifest Schema Main manifest file: - `data/frenchnews7_manifest.csv` Required columns: - `id` - `url` - `publisher` - `label` - `split` - `date` - `year` No text-bearing columns are included. In particular, the release excludes headline, body, snippet, and summary fields. Example rows: ```csv id,url,publisher,label,split,date,year fn7-1089,https://www.lemonde.fr/economie/article/2025/06/23/...,Le Monde,Économie,validation,2025-06-23,2025 fn7-2969,https://www.lemonde.fr/idees/article/2025/06/13/...,Le Monde,Sciences & Technologies,test,2025-06-13,2025 ``` Labels reflect the harmonized benchmark taxonomy and may not correspond literally to the URL path slug in all cases. ## Label Taxonomy - **Société**: domestic social affairs, human interest, crime, health, education, and civil society - **Culture & Loisirs**: arts, cinema, music, television, books, leisure, and lifestyle - **International**: foreign affairs, geopolitics, and global events outside France - **Politique**: domestic politics, government, elections, institutions, and legislative affairs - **Sport**: all sports coverage regardless of discipline - **Économie**: business, finance, macroeconomics, markets, and corporate news - **Sciences & Technologies**: scientific research, technology, digital innovation, and environment-science topics ## Limitations - The release is manifest-only and does not redistribute article text. - Labels reflect a harmonized cross-publisher editorial taxonomy rather than arbitrary fine-grained semantic annotation. - Reconstruction success may vary over time as publisher pages change or disappear. - Performance and coverage may vary across outlets, years, and topic distributions. - Bucket A slug coverage varies by publisher, from 100% (JDD) to 34.3% (Slate.fr). ## Legal Note This repository is designed to support reproducibility without redistributing copyrighted news content. Users are responsible for ensuring that any downstream fetching, storage, processing, or redistribution of source material complies with applicable law, publisher terms, robots directives, and institutional policy. The CC-BY-4.0 license applies only to the original manifest structure, labels, taxonomy definitions, and repository-authored materials released by the authors, and does not apply to the linked publisher content. ## Reconstruction Note The `reconstruction/` folder provides lightweight starter utilities to refetch pages from the original URLs under the user's own legal and operational framework. These utilities are intentionally conservative and do not attempt aggressive scraping or text extraction. They are provided solely as a convenience for authorized users. ## Citation If you use FrenchNews-7 or the accompanying classifier, please cite the associated paper. ```bibtex @misc{sobhy2026frenchnews7, title = {FrenchNews-7: Benchmarking Cross-Publisher French News Editorial Desk Classification}, author = {Amr Sobhy}, year = {2026}, note = {Working paper, version April 2026}, } ``` Project page: