Post
31
Two failure modes we kept hitting in finance documents, and what we published on each:
๐ SAVIOR (WACV 2026 workshop): OCR engines and VLMs still miss rotated text, logo-embedded vendor names, fine print, degraded scans and multi-column layouts. Instead of more data, SAVIOR targets those exact failure cases. A fine-tuned Qwen2.5-VL reaches 0.93 word recall, and 0.89 F1 on document QA with its layout-aware output. It also introduces PaIRS, a metric for layout fidelity.
๐ DocuLite (AAAI 2026 Deployable AI workshop): you can't train extraction models on customer invoices you're not allowed to see. DocuLite generates structure-preserving synthetic data instead, giving +0.525 F1 for a 7B LLM and +0.513 F1 for an 8B VLM on 184 hard real invoices.
Plain-language walkthrough of both: https://huggingface.co/blog/nchhaya/two-papers-on-the-documents-that-break-ocr
Open benchmarks for these failure cases: hyperbots-inc/hyperapi-enterprise-document-intelligence-6af6c8764547d894d4e1a3c4
๐ SAVIOR (WACV 2026 workshop): OCR engines and VLMs still miss rotated text, logo-embedded vendor names, fine print, degraded scans and multi-column layouts. Instead of more data, SAVIOR targets those exact failure cases. A fine-tuned Qwen2.5-VL reaches 0.93 word recall, and 0.89 F1 on document QA with its layout-aware output. It also introduces PaIRS, a metric for layout fidelity.
๐ DocuLite (AAAI 2026 Deployable AI workshop): you can't train extraction models on customer invoices you're not allowed to see. DocuLite generates structure-preserving synthetic data instead, giving +0.525 F1 for a 7B LLM and +0.513 F1 for an 8B VLM on 184 hard real invoices.
Plain-language walkthrough of both: https://huggingface.co/blog/nchhaya/two-papers-on-the-documents-that-break-ocr
Open benchmarks for these failure cases: hyperbots-inc/hyperapi-enterprise-document-intelligence-6af6c8764547d894d4e1a3c4