| ### 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 | |
| Repository: https://github.com/Tobiadefami/docmodel | |
| ### 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 | |
| ``` python | |
| from transformers import AutoTokenizer, AutoModel | |
| tokenizer = AutoTokenizer.from_pretrained("tobiadefami/docmodel-base") | |
| model = AutoModel.from_pretrained("tobiadefami/docmodel-base") | |
| # Example usage | |
| inputs = tokenizer("Your document text here...", return_tensors="pt") | |
| outputs = 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. |