--- library_name: transformers pipeline_tag: token-classification license: cc-by-nc-sa-4.0 base_model: microsoft/layoutlmv3-base language: - vi tags: - layoutlmv3 - kie - key-information-extraction - document-understanding - vietnamese - onnx - int8 --- # LayoutLMv3 — Vietnamese administrative document KIE Fine-tuned [`microsoft/layoutlmv3-base`](https://huggingface.co/microsoft/layoutlmv3-base) for key-information extraction on Vietnamese administrative documents (Quyết định, Công văn, Tờ trình, Báo cáo, ...). The variant is the **fontgray-norm** flavour: `token_type_ids` encode three style buckets derived from per-word font size, foreground gray level, and word height (see `style_emphasis_ids` in the training pipeline). ## Files - `layoutlmv3_fontgray_norm_final_epoch25/layoutlmv3_fontgray_norm_final_epoch25.int8.onnx` — quantized INT8 ONNX model - `layoutlmv3_fontgray_norm_final_epoch25/label_list.json` — label vocabulary (BIO tags) - `layoutlmv3_fontgray_norm_final_epoch25/layoutlmv3_fontgray_config.json` — runtime config (style buckets, line position buckets) - Tokenizer files (`tokenizer.json`, `tokenizer_config.json`, `special_tokens_map.json`, `vocab.txt`, …) ## Intended use This model is designed for the [ScanIndex](https://github.com/welcomyou/scanindex) pipeline. It expects the canonical OCR JSON profile produced by ScreenAI + the project's preprocessing (`layoutlmv3_runtime_v1`). Using it standalone requires reproducing that input format. ## Loading ```python from huggingface_hub import snapshot_download local = snapshot_download("welcomyou/layoutlmv3-vn-admin-kie", local_dir="models") # Then point ScanIndex at /models/layoutlmv3_fontgray_norm_final_epoch25/ ``` ## Training & data See [train-convert/kie/train_kie/layoutlmv3_fontgray_norm/](https://github.com/welcomyou/scanindex/tree/main/train-convert/kie/train_kie/layoutlmv3_fontgray_norm) for the training scripts and decision records. Trained on internal annotated Vietnamese admin documents (not redistributed). ## License Inherits LayoutLMv3 base license: **CC-BY-NC-SA-4.0** (research / non-commercial). Commercial use requires a separate agreement with Microsoft for the base model.