A fine-tuning corpus for Nepali (Devanagari) government-document OCR and layout analysis, built to adapt a vision-language model to Nepali paperwork without erasing what it already knows. Roughly half is the target task; the other half is replay data, mixed in deliberately. That split follows Fine-tuning VLMs Without Forgetting, whose one unresolved failure mode is task-specific overfitting — the model memorises a single output template and stops reading the prompt. Diverse tasks are the measured remedy, so training only on Nepali documents would walk straight into it. pages / crops / organic carry the full Page IR: replay are uniform: image, prompt, target, source, config. Table format is…