Supervised fine-tune of BAGEL-Canvas on the Rush Hour block-sliding puzzle, teaching interleaved visual chain-of-thought: the model reasons step by step and renders the intermediate board states while deciding which car to slide and in which direction.

This checkpoint is the merged model (base + LoRA adapters folded in) at training step 600, ready to load directly — no adapter merge needed.

Load with the BAGEL-Canvas / MathCanvas training-and-inference code. Used as the GRPO starting policy in the Rush Hour RL experiments.