document-vqa

Reproducible framework for document-centric VLM robustness and claim-faithfulness evaluation.

Solutions and technical reports for challenges from the OLP AI PTIT 2026 Preliminary Round.

An enriched multi-page DocVQA benchmark based on DUDE, featuring LLM-judged reading orders, YOLO-detected visual elements, and unanswerable/corruption QA pairs for hallucination mitigation.

Document VQA. A pure-Python rule system that beats the mentor baseline (0.9545 → 0.968), plus measured negative results for LayoutLMv3 and Qwen2.5-VL.