Hugging Face Overview
Chocolatine-2-14B-Instruct-v2.0.3 Overview
Chocolatine-2-14B-Instruct-v2.0.3 is a 14.8 billion parameter instruction-tuned model developed by Jonathan Pacifico, based on the Qwen-2.5-14B architecture. It has been fine-tuned using DPO (Direct Preference Optimization) with the jpacifico/french-orca-dpo-pairs-revised RLHF dataset, specifically enhancing its capabilities in the French language. The model supports a substantial context window of up to 128K tokens.
Key Capabilities & Performance
- French Language Proficiency: Achieves top 3 rankings across all categories on the French Government Leaderboard LLM FR.
- Benchmark Excellence: Recognized as the strongest open-weights model on the COLE Benchmark (Laval University) with a Composite Score of 45.05%. Details are available in this paper.
- MT-Bench-French Performance: Outperforms previous Chocolatine versions and the base Qwen-2.5 model on MT-Bench-French, closely approaching GPT-4o-mini's performance in French.
- Multilingual Support: While primarily optimized for French, it also supports English.
Good For
- Applications requiring high-quality French language generation and understanding.
- Developers seeking a powerful open-source model for French-centric tasks.
- Use cases benefiting from a large context window (up to 128K tokens).
Limitations
This model series serves as a demonstration of effective fine-tuning. It does not include any built-in moderation mechanisms.