Chocolatine-2-4B-Instruct-DPO-v2.1 Overview

Chocolatine-2-4B-Instruct-DPO-v2.1, developed by jpacifico, is a 4.0 billion parameter instruction-tuned model built upon Qwen/Qwen3-4B-Instruct-2507. It features an impressive native context length of 262,144 tokens. The model underwent a multi-stage post-training process involving Direct Preference Optimization (DPO) with French preference data (Compar:IA and French-ORCA pairs) and subsequent model merging using MergeKit with the TIES method.

Key Capabilities

• Enhanced French Performance: Demonstrates consistent gains across various French benchmarks, including gpqa-fr:diamond (32.49), french_bench_arc_challenge (49.79), and global_mmlu_fr (64.75), indicating broad improvements in French language understanding and generation.
• Multilingual Proficiency: While primarily optimized for French, it preserves strong English capabilities, with some tasks showing slight improvements, suggesting positive cross-lingual transfer.
• Instruction Following: Designed to improve instruction-following and reasoning, making it suitable for direct generation tasks.
• Local Inference Optimization: Available in optimized variants like MLX (for Apple silicon) and GGUF (quantized versions like Q4_K_M, Q8_0) for efficient local deployment.

Good For

• French Language Applications: Ideal for use cases requiring high performance in French, such as content generation, question answering, and instruction following in French.
• Resource-Efficient Deployment: Its optimized variants make it well-suited for local inference on consumer hardware.
• Direct Instruction Following: Favors a compact, dense instruct model focused on direct generation efficiency, rather than explicit reasoning traces or structured thinking outputs.