reducto

1 models • 1 total models in database
Sort by:

RolmOCR

Earlier this year, the Allen Institute for AI released olmOCR, an open-source tool that performs document OCR using the Qwen2-VL-7B vision language model (VLM). We were excited to see a high-quality, openly available approach to parsing PDFs and other complex documents — and curious to explore what else might be possible using newer foundation models and some lightweight optimizations. The result is RolmOCR, a drop-in alternative to olmOCR that’s faster, uses less memory, and still performs well on a variety of document types. We're releasing it under Apache 2.0 for anyone to try out, explore, or build on. This model is a fine-tuned version of Qwen/Qwen2.5-VL-7B-Instruct on the full allenai/olmOCR-mix-0225 dataset. 1. New Base Model: We swapped in a more recent version of the existing model (Qwen2.5-VL-7B) as the foundation. 2. No Metadata inputs: Unlike the original, we don’t use metadata extracted from PDFs. This significantly reduces prompt length, which in turn lowers both processing time and VRAM usage — without hurting accuracy in most cases. 3. Rotation of training data: About 15% of the training data was rotated to enhance robustness to off-angle documents. We otherwise use the same training set. - RolmOCR, like other VLM-based OCR solutions, still suffer from hallucination or dropping contents. - Unlike the Reducto Parsing API, RolmOCR cannot output layout bounding boxes. - We have not evaluated the performance of any quantized versions.

NaNK
license:apache-2.0
9,684
562