DocLoom

Performance Comparison of Various Models on the olmOCR-Bench Dataset

Introduction

DocLoom is dedicated to structured text extraction from complex PDFs. Its core goal is to convert unstructured PDF content into linearly organized plain text in natural reading order, while accurately preserving key structural information (e.g., chapter hierarchies, tables, lists, formulas) to provide high-quality data for downstream tasks such as document parsing, information retrieval, and data mining.

1. Based on the Qwen3-VL-4B-Instruct architecture, improve the special adaptability of PDF processing through Supervised Fine-Tuning (SFT).
2. Adopt "existing data + synthetic data" dual-source training, with a focus on enhancing the performance of PDF processing in the medical field.
3. On olmOCR-Bench , compared with other models of the same parameter scale, WiNGPT-OCR-4B has achieved excellent performance.

How to Use

Transformers (Recommended)

Refer to Qwen3-VL-4B-Instruct for guidance on model inference acceleration and PDF processing, etc.

vLLM (Recommended)

For high-performance inference and deployment, we recommend using vLLM. We also provide a standalone script for efficiently processing multi-page PDF documents. This script operates independently and does not require the official olmOCR toolkit, offering a lightweight and fast way to perform OCR on entire documents.

python DocLoom_test.py <pdf_file_path>

Acknowledgement

We express our gratitude to the teams that developed olmOCR and Qwen3-VL, which were instrumental in our research.