KhatianDoc - Document QA EM: leaderboard

KhatianDoc (2026) is a human-verified benchmark on 107 real Bangladeshi RS Khatian land records in a base-16 fractional numeral system, evaluated zero-shot. This board is document question answering scored by exact match on a 300-item stratified subset of 1,634 QA pairs, using the authors refusal-normalized rescoring that credits the prompted refusal sentence where the reference records absent information as a short negation; in percent, higher is better.

Metric: Exact match (%; refusal-normalized, 300-item stratified document-QA subset of 1,634 QA pairs, zero-shot). Source: arxiv.org. Saturation forecast: Around 2032. 6 models tracked.

Top models

#ModelScoreOverall rank
1Gemini 2.5 Flash Lite21.67#958
2Llama 4 Scout18.33#1787
3Qwen 3 VL 8B Instruct13#982
4Qwen 2.5 VL 72B Instruct0.34#845
5Gemma 4 26B A4B (IT)0.33#545
6GPT-4o Mini0#1627

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=khatiandoc-document-qa-em · How It Works · Data refreshed daily, snapshot 2026-11-12.