LipiOCR

The OCR/text-extraction engine behind LipiOCR, a fast document intelligence pipeline (classify → OCR → QR/barcode decode → normalize).

Important: this is a component, not the full pipeline

This model’s weights are identical to glm-ocr — this is not a custom fine-tune. It’s glm-ocr with a document-extraction system prompt baked in as the default. Two things were tried and deliberately not shipped here: fine-tuning this model and a separate model (Granite-Docling-258M) on the LipiOCR document dataset — both collapsed into repetition loops and wrong field extraction on real photographed documents within the first few hundred training steps, despite improving on synthetic validation metrics. Zero-shot performed better than either fine-tune, so that’s what’s published.

The full LipiOCR pipeline includes pieces this model alone does not: - Document type classification (a separate model, not this one) - QR/barcode decoding - Digit-swap correction and field normalization - A two-call extraction strategy (table content, then header/label content, called separately) that avoids this model’s tendency to drop header text or loop when given one compound prompt

If you use this model directly for a document with both a table and surrounding label/value fields, ask for one or the other specifically rather than “extract everything” - a single compound prompt can trigger that dropped-text/repetition behavior.

Usage

Attach a document image and either: - Extract only the data table in this image as an HTML table. - Extract all text from this image, exactly as it appears, one field per line.

Built by

Krishna, Brilliant Systems Solutions Pvt Ltd.