Donut

Use this repo skill when the task involves NAVER/ClovaAI Donut: OCR-free document understanding with a Swin encoder and BART-like decoder, plus SynthDoG synthetic document generation. This skill is self-contained; do not reopen the original checkout for routine Donut usage.

Route by task

User intent	Read or run
Load a Donut checkpoint, run one-image prediction, compare prompts, or start the Gradio demo	sub-skills/inference/SKILL.md
Fine-tune, validate dataset JSONL, evaluate a checkpoint, read metrics, or debug configs/checkpoints	sub-skills/training/SKILL.md
Generate synthetic document images, render SynthDoG configs, or adapt resources/fonts/corpora	sub-skills/synthdog/SKILL.md
Inspect public package APIs, signatures, token helpers, and object relationships	references/api-reference.md
Decide which bundled workflow helper or sub-skill owns a request	references/workflow-map.md
Diagnose install/import, Hugging Face, CUDA, Gradio, dataset, or SynthDoG dependency failures	references/troubleshooting.md
Check source baseline, commit, version, and refresh triggers	references/repo-provenance.md

Fast start

1. Install a Donut-capable Python environment. For inference only, pip install donut-python plus a compatible PyTorch build is the starting point. Training needs CUDA-capable PyTorch. SynthDoG generation also needs synthtiger and its image/text dependencies.
2. Smoke-check the environment without the source repo:

   python scripts/runtime_smoke.py
   python scripts/runtime_smoke.py --check all --require-cuda  # when training and SynthDoG are in scope
   

3. Route to a sub-skill before giving detailed commands. Each sub-skill owns its own scripts and troubleshooting notes.

Core facts to preserve

• Package distribution: donut-python; import module: donut.
• Public exports include DonutConfig, DonutModel, DonutDataset, JSONParseEvaluator, load_json, and save_json.
• DonutModel.from_pretrained(...) loads local or Hugging Face checkpoints and requests the Donut repositories' official revision.
• DonutModel.inference(...) returns a dictionary with a predictions list; parsed JSON is the default output.
• Dataset rows use file_name plus a JSON-encoded ground_truth string containing either gt_parse or gt_parses.
• Training is GPU-oriented in the original scripts; do not promise CPU training equivalence.
• No model weights, datasets, or large SynthDoG assets are bundled in this skill. Users must provide checkpoints, images, datasets, and SynthDoG resources.

Bundled helpers

• scripts/runtime_smoke.py: cross-workflow environment/API smoke check.
• sub-skills/inference/scripts/run_inference.py: single-image inference and prompt comparison.
• sub-skills/inference/scripts/launch_demo.py: Gradio demo launcher.
• sub-skills/training/scripts/train_donut.py: bundled CUDA-oriented trainer entry point.
• sub-skills/training/scripts/check_training_config.py: config and metadata validator.
• sub-skills/training/scripts/evaluate_dataset.py: local/HF dataset evaluation helper and validation-only mode.
• sub-skills/training/references/configs/: bundled CORD, DocVQA, RVL-CDIP, and TrainTicket config examples.
• sub-skills/synthdog/scripts/render_config.py: render bundled SynthDoG placeholder configs for external resources.
• sub-skills/synthdog/scripts/template.py: bundled SynthDoG template used by synthtiger.

Boundaries

Use Donut for OCR-free document-image understanding, classification, extraction, visual QA, and synthetic-document pretraining data generation. Do not use this skill for generic PDF parsing, OCR engines, RAG chunking, non-document computer vision models, or unrelated Transformers training unless Donut APIs, checkpoints, config files, errors, or data formats are central to the request.