ToxiFrench: French Toxicity Benchmark & Methods

• ToxiFrench is an ecosystem of benchmarks, datasets, and methods tailored for detecting and mitigating toxic French text across LLMs and moderation systems.
• It introduces a robust chain-of-thought annotation pipeline combining automated labeling with human verification to ensure high reliability.
• The framework evaluates diverse models and approaches, emphasizing cross-lingual detoxification, dynamic fine-tuning, and scalable safety-critical applications.

TOXIFRENCH refers to the ecosystem of benchmarks, datasets, models, and evaluation methods designed to detect, measure, and mitigate toxic text content in French, with direct application to LLMs, moderation systems, and social research in digital safety. The landscape encompasses pure toxicity detection (classification), generative detoxification (text rewriting), and LLM safety evaluation. The distinctive properties of French—morphology, idiomaticity, sociolinguistic diversity—necessitate specialized datasets and methods that diverge from English-centric paradigms.

1. ToxiFrench Benchmark: Construction, Pipeline, and Dataset Properties

The ToxiFrench benchmark is a corpus of 53,622 French online comments, sourced from high-traffic platforms such as jeuxvideo.com (2011–2025 coverage), processed through a semi-automated annotation pipeline. The construction pipeline comprises two principal stages:

• Pre-annotation via GPT-4o-mini: Every candidate is labeled with a structured chain-of-thought (CoT), a binary label (toxic / non-toxic), and a toxicity score $s_i \in [0,10]$. The auto-label non-toxic rule—$s_i \leq \theta$ with $\theta=3$—exhibits 100%±2% agreement with human annotation on a 422-sample split.
• Human Verification: Applied to 10% of ambiguous or low-confidence cases; intra-annotator agreement is 97–98%, inter-annotator on toxic items is 97%.

This results in a highly scalable, culturally aligned corpus: 52,274 training (4% toxic, 96% non-toxic), and a benchmark set ($S_\text{bench}$) of 1,388 comments, class-balanced (50/50 toxic/non-toxic). The annotation approach systematically reduces manual labor while preserving label reliability (Delaval et al., 15 Aug 2025).

2. Model Benchmarking and Quantitative Results

ToxiFrench provides a unified platform to evaluate over 20 models—large LLMs, small LLMs (SLMs), and transformer classifiers. Key results (all on class-balanced $S_\text{bench}$):

Model	Acc.	macro F1	Prec₀	Rec₀	F1₀	Prec₁	Rec₁	F1₁
GPT-4o (200B)	0.844	0.841	.990	.696	.817	.766	.993	.864
Mistral-Large (123B)	0.857	0.856	.953	.752	.841	.795	.963	.871
GPT-4o-mini (8B)	0.870	0.870	.962	.771	.856	.809	.970	.882

Crucially, SLMs like GPT-4o-mini (8B) outperform larger models, particularly in macro F1 and accuracy, and show greater robustness to subtle, context-dependent cues. Larger models exhibit a bias towards over-predicting toxicity, with high non-toxic precision but lower recall for toxic content (Delaval et al., 15 Aug 2025).

3. Chain-of-Thought (CoT) Fine-Tuning for Toxicity Detection

ToxiFrench introduces a CoT fine-tuning regime with dynamic weighted loss. During supervised learning, each sample includes both a rationale $r = [r_1,\ldots, r_n]$ within > … and a final binary label $y$.

Loss decomposition: $$\mathcal{L}^{(e)} = \lambda^{(e)}_{r}\,\mathcal{L}_{r}+\lambda^{(e)}_{y}\,\mathcal{L}_{y}$$ where $\mathcal{L}_{r}$ is cross-entropy on the rationale and $\mathcal{L}_{y}$ on the final answer; weights are updated geometrically each epoch to increasingly emphasize $s_i \leq \theta$0.

Performance gain: rec-SOAP model (Qwen-3-4B, CoT + dynamic loss) achieves macro F1 = 0.855, outperforming GPT-4o and Gemini-2.5-Flash. Zero-shot F1 improves from 0.728 to 0.857 with CoT fine-tuning (Delaval et al., 15 Aug 2025).

This approach promotes “faithfulness”—the final binary decision aligns with the model’s explicit rationale—allowing for more reliable moderation and increased user trust.

4. Detoxification: Generation, Evaluation, and Trade-Offs

Detoxification in French generative models is addressed through lexicon-guided and classifier-gated rewriting (Lai-Lopez et al., 24 Jul 2025), and cross-lingual transfer methodologies (Dementieva et al., 2023):

• Lexicon-Guided Tagging: Toxic spans in French are pre-tagged using the multilingual_toxic_lexicon, with e.g., <toxic>connard</toxic>, before model input.
• Sequence-to-Sequence Model: Instruction-tuned mT5-XL is fine-tuned on both human-curated and synthetic toxic→neutral French pairs.
• Classifier-Gated Iteration: A multilingual DistilBERT classifier assesses output toxicity probability; if above $s_i \leq \theta$1, the generator is rerun (up to 2 passes).

Key French results ($s_i \leq \theta$2, $s_i \leq \theta$3, $s_i \leq \theta$4, $s_i \leq \theta$5) reveal a trade-off: increasing style transfer accuracy (removal of toxicity) can reduce semantic similarity. Conservative thresholds preserve meaning and fluency but leave residual toxic content. Example output:

• Input: « Ferme ta gueule, connard. »
• Output: « Veuillez vous taire, s’il vous plaît. »

Recommendations include continual lexicon updates, augmentation with synthetic hard cases, and threshold tuning per QA balance (Lai-Lopez et al., 24 Jul 2025).

5. Cross-Lingual and Data-Efficient Approaches

Methodologies for French detoxification in low-resource or cross-lingual scenarios are distilled as:

• Adapter-based Transfer: Fine-tune lightweight adapters inside multilingual LMs (mBART, mT5) on English ParaDetox, supplemented by French paraphrase/translation corpora. This yields strong content fidelity and fluency with no need for end-to-end NMT in deployment (Dementieva et al., 2023).
• Metrics for Evaluation: Post-detox outputs are scored via French-specific classifiers (STA), semantic similarity (BLEURT_FR or paraphrase classifier; SIM), and grammatical acceptability (CoLA-French; FL). Joint score $s_i \leq \theta$6 aggregates these signals and correlates more strongly with human judgment than prior metrics.

Challenges remain, including robustness to slang/misspelling, risk of semantic loss in over-detoxification, and scarcity of parallel French detox corpora.

6. Large-Scale Prompt-Based Toxicity Evaluation in French LLMs

The FrenchToxicityPrompts benchmark (Brun et al., 2024) evaluates generative toxicity across LLMs using 50K French Reddit prompt–continuation pairs, scored by the multilingual Perspective API on six axes (toxicity, severe_toxicity, identity_attack, insult, profanity, threat).

Key metrics:

• Expected Maximum Toxicity (EMT)
• Toxicity Probability (TP): probability that any of 25 continuations exceeds $s_i \leq \theta$7
• Toxic Fraction (TF)
• Average Toxicity (AT)

Model and prompt effects:

Model family	Size	EMT (toxic/non-toxic)	TP	TF	AT
BLOOMZ (7.1B, instr.)	7.1B	0.50/0.30	0.15/0.06	0.07/0.02	0.12/0.05

Model size increases worst-case toxicity, but instruction-tuned models (BLOOMZ, LLaMa2-chat) exhibit substantially lower rates. Toxic prompts disproportionately inflate generated toxicity, necessitating pre-filtering for safety.

Recommended interventions:

• Prompt filtering before generation (e.g., Detoxify)
• Model selection favoring instruction-tuned or data-governed architectures
• Decoding-time controls (DExperts), ongoing live monitoring, and dialectal/idiomatic robustness for French (Brun et al., 2024).

7. Limitations, Gaps, and Future Directions

While the ecosystem around ToxiFrench has advanced both detection and mitigation, several limitations persist:

• No direct cross-lingual detoxification analysis for French in “Breaking mBad!” (Beniwal et al., 22 May 2025); transfer results from neighboring languages are only suggestive.
• Dataset coverage for idiomatic, regional, or low-intensity French toxicity remains incomplete.
• Epistemological challenges in annotation and model calibration for subtle pragmatic toxicity are only partially addressed by CoT and classifier-based methods.

Anticipated directions include the expansion of pipeline annotation to more languages, curriculum learning for “faithfulness,” and multimodal toxicity detection incorporating social context.

---

The ToxiFrench suite establishes the technical and infrastructural baseline for scalable, robust French toxicity research, spanning detection, mitigation, benchmarking, and policy-relevant deployment scenarios. It synthesizes advances in LLM fine-tuning, classifier-based evaluation, and large-scale, culturally situated corpus construction for maximal impact in safety-critical applications (Delaval et al., 15 Aug 2025, Lai-Lopez et al., 24 Jul 2025, Dementieva et al., 2023, Brun et al., 2024, Beniwal et al., 22 May 2025).