LayoutReader-base

Detect the reading order of a layout-rich document (e.g. receipts, forms, papers) by encoding the word-level bounding boxes in a seq2seq transformer.

Refer to the official Github Repo: https://github.com/microsoft/unilm/tree/master/layoutreader

Model details

Trained on ReadingBank.

Architecture	LayoutlmForSequenceToSequence (LayoutLM backbone)
Layers / hidden / heads	12 / 768 / 12
Parameters	~113M (207 tensors, fp32)
Vocab size	30522
Max position embeddings	1024
Max 2D position embeddings	1024
Max source length	513

Files

• model.safetensors / pytorch_model.bin — the same checkpoint in both formats
• config.json

No tokenizer files are included in this repo — load one separately.

Note on revisions

The original upload of this repo was saved through a plain BertModel, which silently dropped every LayoutReader-specific tensor: the 2D layout position embeddings (x/y/h/w_position_embeddings) and the cls.predictions head. Its config.json also recorded architectures: ["BertModel"].

main now holds the complete original checkpoint, verified against the source pytorch_model.bin.

The previous incomplete version is still available at the v1-old tag:

from transformers import AutoModel
AutoModel.from_pretrained("zilongwang/layoutreader-base", revision="v1-old")