💬 NLP

ToxiFrench: Benchmarking and Enhancing Language Models via CoT Fine-Tuning for French Toxicity Detection

This paper introduces ToxiFrench, a large-scale French toxicity dataset created via a semi-automated pipeline, and demonstrates that a 4B parameter model fine-tuned with a novel Chain-of-Thought strategy and Dynamic Weighted Loss outperforms larger models and state-of-the-art systems in detecting French toxic content.

Original authors: Axel Delaval, Shujian Yang, Haicheng Wang, Han Qiu, Jialiang Lu

Published 2026-04-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Axel Delaval, Shujian Yang, Haicheng Wang, Han Qiu, Jialiang Lu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to spot rude, mean, or dangerous comments on a French internet forum. This is the challenge the authors of this paper tackled. They didn't just throw a big, expensive computer at the problem; they built a smarter, more efficient system.

Here is the story of TOXIFRENCH, broken down into simple parts with some creative analogies.

1. The Problem: The "Lost in Translation" Gap

For years, researchers have been great at teaching computers to spot toxic English comments. But French is different. It's like trying to teach someone to recognize a joke in a language where the punchline depends on a specific cultural reference or a clever play on words.

The Issue: Existing French datasets were either just English comments translated (which feels fake and misses the cultural nuance) or too small to be useful.
The Result: Computers were either missing subtle French insults or flagging innocent French jokes as dangerous.

2. The Solution: Building a Better "Training Ground"

The team created TOXIFRENCH, a massive library of 53,000 real French comments. But they didn't just ask humans to read them all (that would take forever and be emotionally draining).

The Hybrid Approach: They used a "Smart Assistant" (a large AI) to do the heavy lifting. The AI read the comments, wrote a short explanation of why it thought they were toxic, and gave a score.
The Human Safety Net: Humans only stepped in to check the "tricky" 10% of cases where the AI wasn't sure.
The Analogy: Imagine a senior teacher (the Human) grading a stack of 1,000 essays. Instead of reading every single one, they let a very smart teaching assistant (the AI) grade 900 of them. The assistant writes a note explaining its grade. The teacher only reads the 100 essays where the assistant was unsure or where the grade seemed weird. This saves time but keeps the quality high.

3. The Big Surprise: The "Small Dog" vs. The "Huge Elephant"

Usually, in AI, bigger is better. A giant model with billions of parameters is expected to beat a smaller one. But the researchers found something surprising:

The Finding: A relatively small model (called Qwen3-4B) actually did a better job at spotting French toxicity than massive, famous models like GPT-4o.
The Analogy: Think of a giant, slow-moving elephant (the big AI) trying to navigate a dense, narrow forest. It might get stuck or miss small paths. Now, imagine a nimble, small dog (the small AI). It can dart through the trees, sniff out the specific smells of the forest, and find the bad actors much faster and more accurately. The small model was more "culturally agile."

4. The Secret Sauce: "Chain-of-Thought" with a Twist

To make that small model even better, they taught it a new way of thinking called Chain-of-Thought (CoT).

How it works: Instead of just shouting "TOXIC!" or "NOT TOXIC!", the model is forced to write a step-by-step reasoning process first. It has to explain why it thinks something is mean before it gives the final answer.
The Innovation: They added a special "Dynamic Weighted Loss."
- The Analogy: Imagine a student taking a test. Usually, the teacher grades the whole answer sheet equally. But here, the teacher says, "I don't care as much about your long, rambling explanation as I do about your final answer."
- As the model trains, the teacher gradually stops caring about the reasoning steps and starts caring only about getting the final "Yes/No" right. This forces the model to use its reasoning to actually get the right answer, rather than just rambling.

5. The Result: A Super-Small, Super-Strong Guard

By combining the new dataset, the small model, and this special training method, they created a toxicity detector that:

Outperforms giants: It beats models that are 10x or 20x larger.
Saves energy: It uses much less electricity (like a hybrid car vs. a gas-guzzling truck).
Is culturally smart: It understands French sarcasm, slang, and hidden insults better than the big models.
Is surprisingly versatile: Even though it was trained only on French, it can spot toxicity in other languages (like German or Chinese) because it learned the logic of hate, not just the specific words.

Summary

The paper is a story about efficiency and cultural understanding. It proves that you don't need a "super-computer" to solve complex social problems. Instead, you need the right data, a smart training method that forces the AI to think before it speaks, and a model that is small enough to be agile but smart enough to understand the nuances of human culture.

They built a smart, nimble guard dog that can sniff out French toxicity better than the giant, clumsy elephants currently patrolling the internet.

1. Problem Statement

The paper addresses the critical gap in French toxicity detection compared to English and Chinese. While significant progress has been made in English, French remains underdeveloped due to:

Lack of High-Quality Data: Existing resources are either small-scale, translated from English (lacking cultural nuance), or rely on automated classifiers (e.g., Perspective API) that fail to capture French-specific linguistic complexities.
Linguistic Complexity: French possesses high orthographic density (e.g., the sound [e] can be written as é, è, ai, er, ez, etc.), enabling coded toxicity and phonetic masking that token-based models often miss.
Cultural Nuance: French toxicity often relies on sarcasm, euphemisms, and context-specific insults that require deep cultural understanding rather than simple keyword matching.
Model Efficiency: Large Language Models (LLMs) are computationally expensive and ecologically costly, raising questions about whether smaller models can achieve comparable safety performance.

2. Methodology

A. Dataset Construction: TOXIFRENCH

The authors released TOXIFRENCH, a dataset of 53,622 native French online comments.

Source: High-traffic French forums (specifically jeuxvideo.com), covering a temporal span from 2011 to 2025 to ensure robustness against evolving trends.
Annotation Pipeline (Semi-Automated):
- LLM Pre-annotation: GPT-4o-mini was used to generate structured Chain-of-Thought (CoT) reasoning for all samples. The model reformulated the text, assessed tone/intent, categorized implicit toxicity (e.g., dog-whistles, irony), and assigned a toxicity score (0–10).
- Human Verification: To ensure quality, a high-confidence rule was applied. If the LLM predicted "non-toxic" or assigned a score $\le 3$ , the label was accepted automatically (covering ~90% of the data). The remaining ~10% (ambiguous/edge cases) were manually annotated by native French speakers.
- Quality Control: The pipeline achieved 97–98% intra-annotator agreement and 90% inter-annotator agreement, establishing a human performance ceiling of ~90% for this task.
Definition of Toxicity: A broad, culturally grounded definition focusing on content causing emotional distress, including hate speech, dehumanization, and threats, but also subtle forms like dog-whistles and implicit bias.

B. Benchmarking Strategy

The authors evaluated a diverse range of models on a balanced test split ( $N=1,388$ ):

Models Tested: State-of-the-art LLMs (GPT-4o, DeepSeek-R1, Gemini-2.5), Small Language Models (SLMs like Qwen3-4B, Mistral-7B), BERT-based encoders (CamemBERT), and commercial APIs.
Key Finding: Contrary to the assumption that larger models are always better, SLMs often surpassed larger LLMs in robustness and generalization on this specific task.

C. Proposed Enhancement: CoT Fine-Tuning with Dynamic Weighted Loss (DWL)

To improve SLM performance, the authors proposed a novel fine-tuning framework using Qwen3-4B:

Chain-of-Thought (CoT) Training: The model is trained to output a structured reasoning process (reformulation, tone analysis, intent assessment) before the final binary decision.
Dynamic Weighted Loss (DWL): Standard SFT treats all tokens equally, diluting the importance of the final answer within a long CoT sequence. The authors introduced a loss function where the weight of the final classification token ( $\alpha_{answer}$ ) progressively increases over training epochs, while the weight of reasoning tokens ( $\alpha_{think}$ ) decreases. This forces the model to master the reasoning format early but prioritize faithfulness to the final decision later.
Optimization: The framework combines Oversampling (to handle class imbalance), SOAP optimizer (for stability), and Direct Preference Optimization (DPO) using synthetic preference pairs generated from the training set.

3. Key Contributions

TOXIFRENCH Dataset: The largest high-quality, culturally grounded public French toxicity dataset (53k+ samples) capturing both overt and subtle toxicity, constructed via a rigorous semi-automated pipeline.
SLM Superiority Finding: Demonstrated that Small Language Models (specifically Qwen3-4B) can outperform much larger models (like GPT-4o and Gemini-2.5) in French toxicity detection when properly fine-tuned.
Novel Training Framework: Introduced a CoT fine-tuning strategy with Dynamic Weighted Loss (DWL). This approach significantly improves model "faithfulness" (alignment between reasoning and final decision) without requiring additional data or massive compute resources.
Cross-Lingual Generalization: Showed that a model fine-tuned exclusively on French data can effectively detect toxicity in other languages (German, Chinese, Italian, Spanish) by internalizing a language-agnostic "logic of toxicity."

4. Results

Performance: The fine-tuned Qwen3-4B model achieved a balanced accuracy of 87% on the TOXIFRENCH benchmark.
- This represents a 10% improvement over its baseline.
- It outperformed GPT-4o (84%) and DeepSeek-R1 (84.2%) on this specific benchmark.
- It surpassed the best-performing BERT-based model (CamemBERT) by a significant margin.
Ablation Study:
- Oversampling was crucial to overcome class imbalance.
- DWL was the primary driver of performance gains, ensuring the model didn't just "hallucinate" reasoning but aligned its conclusion with the logic.
- DPO provided a final boost, refining the model's alignment with human preferences.
Robustness: The model maintained high performance on external benchmarks (RTP-LX, Jigsaw) and demonstrated strong zero-shot capabilities across multiple languages, despite being trained only on French.

5. Significance and Implications

Efficiency vs. Performance: The work challenges the "bigger is better" paradigm in safety alignment, proving that Small Language Models (SLMs) can be more robust and ecologically efficient for specific safety tasks when equipped with the right reasoning frameworks.
Cultural Specificity: It highlights the necessity of native, culturally specific datasets for non-English languages. Translated datasets fail to capture the nuances of French toxicity (e.g., phonetic masking, specific slang).
Scalable Safety: The proposed CoT + DWL framework offers a scalable blueprint for deploying reliable, low-cost toxicity detection systems on edge devices or in resource-constrained environments.
Faithfulness in AI: By prioritizing the final decision through dynamic weighting, the paper addresses the issue of "unfaithful" reasoning in LLMs, where the model generates plausible-sounding but incorrect justifications.

In conclusion, TOXIFRENCH provides a foundational resource for French NLP safety and demonstrates that combining culturally grounded data with structured reasoning fine-tuning allows smaller models to achieve state-of-the-art performance, surpassing even the largest commercial LLMs in specific safety domains.