Stockmark-Nemotron-3-Nano-Omni-JapanDocReader for Japanese Document Parsing

This model adapts a 30B reasoning multimodal model for Japanese structured document parsing using capability injection and forgetting control. Experiments show that mixed SFT prevents the VQA forgetting observed during parsing-centric fine-tuning, while DAPO-based RL further optimizes structured parsing performance.