kurakurai/Luth-LFM2-700M

Luth-LFM2-700M: Enhanced French Capabilities for Small Language Models

Luth-LFM2-700M is a 0.7 billion parameter model developed by kurakurai in collaboration with Liquid AI. It is a fine-tuned version of LFM2-700M, specifically optimized for French language tasks. The model was trained using full fine-tuning on the Luth-SFT dataset with Axolotl, a process that successfully improved its French performance while retaining its English capabilities.

Key Capabilities and Performance

This model demonstrates significant improvements in French benchmarks, including:

• Instruction Following (IFEval French): Achieves 50.22, outperforming LFM2-700M (41.96) and other comparable models like Llama-3.2-1B (27.79) and Qwen3-0.6B (44.86).
• General Knowledge (GPQA-Diamond French, MMLU French): Scores 27.92 and 44.72 respectively, showing notable gains over its base model and other small LLMs.
• Mathematics (Math500 French): Reaches 38.40, a substantial improvement from LFM2-700M's 32.40.

Crucially, its English benchmark scores remain stable, with strong performance in IFEval English (63.40) and MMLU English (50.39), indicating effective cross-lingual transfer without degradation.

Ideal Use Cases

• French-centric applications: Excels in tasks requiring strong French instruction following, mathematical reasoning, and general knowledge.
• Bilingual systems: Suitable for applications that need to operate effectively in both French and English, leveraging its retained English capabilities.
• Resource-constrained environments: As a 0.7B parameter model, it offers efficient performance for deployment where larger models are impractical.