ToxiFrench: 53,622 French Comments for Toxicity Detection (2011–2025)

53,622 native French comments collected between 2011 and 2025 address a gap in non-English toxicity detection. The dataset, created by Axel Delaval of École Polytechnique and Shanghai Jiao Tong University, is designed to improve models' understanding of cultural nuances and coded toxicity in French.

Use Cases

• Training French toxicity classifiers based on the large-scale collection of native comments.
• Fine-tuning language models for safety and moderation tasks based on the annotated French text.
• Researching cultural and coded toxicity in French online discourse based on the dataset's focus on nuance.
• Developing DPO (Direct Preference Optimization) or Chain of Thought models for safety alignment using French examples.

Strengths

• Contains 53,622 examples, providing a substantial corpus for model training.
• Covers a 14-year time range (2011–2025), potentially capturing evolving language and toxicity patterns.
• Focuses on native French comments, which may improve model performance on cultural nuance.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but specific annotation details, label distribution, and data quality require manual inspection.
• Data may reflect temporal and source bias inherent to the collection method, which is not described.

Provenance

Source

Axel Delaval (AxelDlv00) via Hugging Face.

Collection Method

Collection method is not specified in the provided description.

Time Range

2011–2025

Freshness

Last updated 2026-01-02 13:25:41; freshness should be verified.

Geography

null