LoRA adapter for Qwen/Qwen3-4B-Instruct-2507, fine-tuned on the vi-gsm8k-agentic dataset — native Vietnamese grade-school math generated by an agentic self-instruct pipeline (beats machine-translated data).

Multi-seed (3 seeds): in-dist 80.8% vs translated baseline 76.5% (+4.3); OOD 90.2% vs 83.6% (+6.6).