Luth-1.7B-Instruct: Enhanced French Capabilities

Luth-1.7B-Instruct is a 2 billion parameter instruction-tuned model developed by kurakurai, based on the Qwen3-1.7B architecture. It has been specifically fine-tuned on the Luth-SFT dataset to drastically improve its performance in French across various tasks, including instruction following, mathematical reasoning, and general knowledge.

Key Capabilities and Performance

• Bilingual Proficiency: While primarily focused on French enhancement, the model successfully retains and even improves its English capabilities in certain areas.
• Instruction Following: Demonstrates strong instruction adherence in both French and English contexts.
• Benchmark Improvements: Achieves notable gains in French benchmarks:
  • IFEval French: 58.53 (vs. 54.71 for Qwen3-1.7B)
  • GPQA-Diamond French: 36.55 (vs. 31.98 for Qwen3-1.7B)
  • MMLU French: 49.75 (vs. 28.49 for Qwen3-1.7B)
  • Math500 French: 62.60 (vs. 60.40 for Qwen3-1.7B)
• Training Methodology: Utilizes full fine-tuning with Axolotl, merging the fine-tuned model back with the base Qwen3-1.7B to preserve original strengths.

Good For

• Applications requiring strong French language understanding and generation.
• Tasks involving instruction following, math, and general knowledge in French.
• Use cases where maintaining solid English performance alongside enhanced French capabilities is crucial.