Carbon-X completed supervised fine-tuning across 3 epochs over 20 dataset subsets, followed by two rounds of asynchronous GRPO reinforcement learning