LayoutLMv3 Japanese Preview
LayoutLMv3 Japanese Preview adapts the LayoutLMv3 architecture to Japanese documents. Its tokenizer comes from Waseda's Japanese RoBERTa, and its visual tokenizer uses BEiT v2 VQ-KD.
Preview
The processor needs Japanese word alignment
The standard LayoutLMv3Processor cannot be used unchanged. The card combines the Japanese tokenizer with LayoutLMv3ImageProcessor and aligns word boxes to subwords. PyMuPDF is an optional PDF-rendering and preprocessing tool, not an included learned OCR model.LayoutLMv3 Japanese Preview official model card
Check OCR and bounding boxes first
Practical guidance. A document workflow needs reliable Japanese text and word positions before this encoder can use them. Evaluate extraction quality through that full input path instead of attributing every error to the model.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Preview |
| Access | Downloadable |
| Model identifiers | layoutlmv3-japanese-preview |
| Training data | More than 20 million NDL WARP pages, plus PubLayNet and DocLayNet.LayoutLMv3 Japanese Preview official model card |
| License | Apache 2.0 for the published model.LayoutLMv3 Japanese Preview official model card |
| Model role | Pretrained document encoder; task heads require downstream fine-tuning.LayoutLMv3 Japanese Preview official model card |
Related models
- LLM-jp ModernBERT Base
LLM-jp / NII LLMC · LLM-jp ModernBERT
A Japanese pretrained text encoder with an 8,192-token sequence limit and ModernBERT architecture.
Launch date not confirmedResearch releaseDownloadable - LLM-jp-4 VL 9B
LLM-jp / NII LLMC · LLM-jp-4 VL
A Japanese-English vision-language reasoning model that changes both the language backbone and training data from the beta.
Launch date not confirmedResearch releaseDownloadable