RomDev2/layoutlm-camembertv2

MODEL_NAME

This repository contains layoutlm-camembertv2 weights exported to safetensors format.

Source

These weights are derived from pretrained models:

• Layout encoder (LayoutLM): microsoft/layoutlm-base-uncased — pretrained on IIT-CDIP + masked visual-language modeling (LayoutLM paper)
• Text encoder: almanach/camembertv2-base — French language model (RoBERTa-like architecture)

Methodology

This checkpoint was produced by weight merging, not end-to-end training.

1. Load the pretrained layout encoder weights (LiLT or LayoutLM) — kept intact
2. Replace the text encoder weights (embeddings, attention layers, FFN) with those from the French model
3. Update the tokenizer and vocabulary configuration accordingly

No training or fine-tuning was performed at this stage.
This checkpoint is intended as a starting point for downstream fine-tuning on French document understanding tasks (NER, token classification, extractive QA…).

Files

File	Description
model.safetensors	Model weights
pytorch_model.bin	Model weights (PyTorch format)
config.json	Model configuration
tokenizer_config.json	Tokenizer configuration
README.md	This model card

Usage

from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("USERNAME/MODEL_NAME")
model     = AutoModel.from_pretrained("USERNAME/MODEL_NAME")

Limitations

• This model has not been fine-tuned on any French document dataset
• Performance on downstream tasks is not guaranteed without task-specific fine-tuning
• Intended for research and experimentation purposes

License

Weights are derived from models released under the MIT and Apache-2.0 licenses.
Please refer to the original repositories for full license terms.

Acknowledgements

• LayoutLM: Pre-training of Text and Layout for Document Image Understanding — Xu et al., 2020
• microsoft/layoutlm-base-uncased

Note: This is not an official release from any of the above organizations.