Model Description
DocModel is a document understanding model built on the RoBERTa architecture. It captures both textual content and 2D spatial relationships, making it ideal for tasks that require processing complex document layouts, such as forms, tables, and scanned documents.
Developed by: Oluwatobi Adefami, Madison May
Model type: Document Understanding (Information Extraction)
License: Apache-2.0
Model Sources
Uses
DocModel can be directly used for document processing, form understanding, and entity extraction from structured and semi-structured documents.
Out-of-Scope Use
Not recommended for tasks that involve purely textual data without layout components or heavily distorted document scans.
Bias, Risks, and Limitations
DocModel’s performance may degrade on highly noisy or poorly structured documents, such as extreme distortions or low-resolution scans.
Recommendations
Users should be mindful of the model’s limitations, particularly in handling documents with severe layout inconsistencies.
How to Get Started with the Model
1from transformers import AutoTokenizer, AutoModel
2
3tokenizer = AutoTokenizer.from_pretrained("tobiadefami/docmodel-base")
4model = AutoModel.from_pretrained("tobiadefami/docmodel-base")
5
6# Example usage
7inputs = tokenizer("Your document text here...", return_tensors="pt")
8outputs = model(**inputs)
Evaluation
Metrics
Eval Loss: 1.36752
F1-Score: 0.84126
Results
DocModel has been evaluated on the FUNSD dataset for information extraction tasks, demonstrating competitive performance in both loss and F1-score.
