This paper presents TOXIFRENCH, a new public benchmark to address the challenges of detecting toxic content in French. TOXIFRENCH consists of 53,622 French online comments and was generated through a semi-automated annotation pipeline that reduces manual labeling to 10% through LLM-based pre-annotation and human validation. Benchmarking various models revealed that small language models (SLMs) outperform large models in terms of robustness and generalization on toxicity detection tasks. Building on this, we propose a novel Chain-of-Thought (CoT) fine-tuning strategy that utilizes a dynamic weighted loss that progressively emphasizes the model's final decision. The fine-tuned 4B model achieves state-of-the-art performance, outperforming LLMs such as GPT-40 and Gemini-2.5. Furthermore, TOXIFRENCH demonstrates robust multilingual capabilities on cross-lingual toxicity benchmarks, suggesting that our methodology can be effectively extended to other languages and safety-critical classification tasks.
