| --- |
| library_name: transformers |
| pipeline_tag: image-to-text |
| base_model: naver-clova-ix/donut-base |
| tags: |
| - donut |
| - document-parsing |
| - lhdn |
| - ebc |
| --- |
| |
| # LHDN Document Parser · Donut · Development |
|
|
| A Donut-based document-image parser for the acceler-technology team's LHDN document-processing project. This repository holds the development checkpoint for the team's model training and iteration work. |
|
|
| **中文简介：** 团队用于 LHDN 文档图片解析的 Donut 开发模型，探索从页面图片生成文档内容的文本表示。 |
|
|
| ## Purpose and model details |
|
|
| Donut combines a visual encoder with a text decoder. It reads a document image and generates an output sequence without requiring a separate OCR model in the Donut architecture. PDF pages must first be rendered as images. |
|
|
| | Item | Value | |
| | --- | --- | |
| | Role | Development checkpoint | |
| | Architecture | `VisionEncoderDecoderModel` | |
| | Encoder / decoder | Donut Swin / mBART | |
| | Recorded base | `naver-clova-ix/donut-base` | |
| | Encoder image-size setting | `[1800, 1350]` | |
| | Saved Transformers version | 4.47.0 | |
|
|
| Source: [config.json](https://huggingface.co/acceler-technology/dev-lhdn-parse-donut-model/blob/main/config.json). Development and production snapshots have different image-size settings. Use this repository's processor alongside its weights. |
|
|
| ## Load the model |
|
|
| Install `torch`, `transformers`, `sentencepiece`, and `pillow` first. |
|
|
| ```python |
| from transformers import DonutProcessor, VisionEncoderDecoderModel |
| |
| model_id = "acceler-technology/dev-lhdn-parse-donut-model" |
| processor = DonutProcessor.from_pretrained(model_id) |
| model = VisionEncoderDecoderModel.from_pretrained(model_id) |
| model.eval() |
| ``` |
|
|
| This is a loading example. The task prompt, image preparation, and output decoding must match the training setup. It is not a validated extraction demo. |
|
|
| ## Project and limitations |
|
|
| Related project data is held in [lhdn-dataset](https://huggingface.co/datasets/acceler-technology/lhdn-dataset), which remains private. The model history records training iterations. The exact training subset, full recipe, and quantitative evaluation results are not documented here. |
|
|
| Poor scans and unfamiliar layouts can affect generated output. Check extracted values against the source document before use. The model does not determine tax or legal outcomes. |
|
|
| ## Team and links |
|
|
| Developed within [acceler-technology](https://huggingface.co/acceler-technology). Henry ([HenryTeo27](https://huggingface.co/HenryTeo27)) led this project area, executed the data work, and participated in model training and technical discussions. [kwyong](https://huggingface.co/kwyong) handled model uploads. |
|
|
| [Project showcase](https://github.com/HenryTeo27/EBC-HuggingFace-Showcase) · [Production Donut snapshot](https://huggingface.co/acceler-technology/prod-lhdn-parse-donut-model) |
|
|
| A license has not been specified for this checkpoint. Check the checkpoint's and upstream model's terms before reuse. |
|
|
