reducto/RolmOCR
RolmOCR by Reducto AI is a 7 billion parameter vision language model (VLM) fine-tuned from Qwen2.5-VL-7B-Instruct, specifically optimized for document Optical Character Recognition (OCR). It serves as a faster and more memory-efficient alternative to olmOCR, designed for parsing PDFs and complex documents. This model excels at extracting plain text from various document types, including those with off-angle orientations, by streamlining input processing.
Loading preview...
RolmOCR: Faster, Lighter Open Source OCR
RolmOCR, developed by Reducto AI, is a 7 billion parameter vision language model (VLM) designed for efficient Optical Character Recognition (OCR) of documents. It is a fine-tuned version of the Qwen/Qwen2.5-VL-7B-Instruct model, specifically trained on the allenai/olmOCR-mix-0225 dataset. This model aims to provide a high-quality, open-source solution for parsing PDFs and other complex documents, offering improvements over its predecessor, olmOCR.
Key Capabilities & Differentiators
- Enhanced Performance: RolmOCR is engineered to be faster and use less memory compared to the original olmOCR, making it suitable for more resource-constrained environments.
- Streamlined Input: Unlike olmOCR, it does not require metadata inputs from PDFs, significantly reducing prompt length, processing time, and VRAM usage without compromising accuracy for most cases.
- Robustness to Rotation: Approximately 15% of its training data was rotated, enhancing its ability to accurately process off-angle documents.
- Vision-Language Model: As a VLM, it processes image inputs directly to extract text, making it ideal for document understanding tasks.
Use Cases & Limitations
RolmOCR is primarily designed for extracting the plain text representation of documents from images. Developers can host it using vLLM and integrate it via an OpenAI-compatible API. While effective for OCR, users should be aware of common VLM limitations such as potential hallucinations or dropped content. Additionally, RolmOCR does not provide layout bounding boxes, a feature available in Reducto's Parsing API.