donut

Why we included this project

Most document parsers chain together separate OCR steps, first detecting and recognizing text, then feeding the results into a layout or extraction model. Donut skips that pipeline: it is a single transformer that takes a document image and outputs structured data directly, trained end to end without any external OCR engine. The repo ships pre-trained checkpoints for document classification, parsing, and document VQA, so a small team can get a working parser running without standing up OCR services of their own. There is also SynthDoG, a companion generator that produces realistic training documents in English, Chinese, Japanese, and Korean, which helps when you have little annotated data. The model is integrated into Hugging Face Transformers, which keeps deployment and fine-tuning within a familiar toolchain.