lilt-only-base

Layout-only pretrained checkpoint from the official LiLT repository.

This is not a complete model — it contains only the 2D spatial (layout) encoder, with no text encoder. It is intended as a building block for combining with any RoBERTa-like text encoder.

What is this?

LiLT (Language-Independent Layout Transformer) decouples text and layout understanding into two separate encoders. lilt-only-base contains exclusively the layout encoder weights, pretrained on document layout understanding (IIT-CDIP dataset).

This allows combining it with any RoBERTa-compatible text encoder to produce a language-specific document understanding model.

Usage

Use gen_weight_roberta_like.py from the official repository to combine with your text encoder of choice:

python gen_weight_roberta_like.py \
     --lilt lilt-only-base/pytorch_model.bin \
     --text your-roberta-model/pytorch_model.bin \
     --config your-roberta-model/config.json \
     --out lilt-your-language-base

Compatible text encoders: any RoBERTa-like model (roberta-base, camembert-base, microsoft/infoxlm-base, etc.)

Files

File	Description
model.safetensors	Layout encoder weights (safetensors format)
pytorch_model.bin	Layout encoder weights (PyTorch format)
config.json	Model configuration (model_type: liltrobertalike)

Note on model type

This checkpoint uses model_type = liltrobertalike, a custom type defined in the original LiLT repository. It cannot be loaded directly with AutoModel from HuggingFace transformers without first combining it with a text encoder via the procedure above.

License

MIT — following the original jpwang/lilt repository.

Acknowledgements

• LiLT: A Simple yet Effective Language-Independent Layout Transformer for Structured Document Understanding — Wang et al., 2022
• Original weights: jpwang/lilt

Note: This is not an official HuggingFace release from the original authors.