Systematic SFT, GRPO for Qwen3-4B. We explore diverse dataset compositions and training recipes to benchmark and improve performance across tasks.