This human-verified ground-truth dataset and evaluation benchmark is built from rare, early 20th-century Devanagari (Hindi) printed books. It provides dual-adjudicated transcriptions and dedicated held-out evaluation sets for assessing OCR and vision-language model (VLM) recognition accuracy on challenging historical texts.
Potential Use Cases
- Historical Hindi OCR/VLM Benchmarking: Cross-verified ground-truth transcriptions can be used to evaluate the character recognition accuracy of OCR and VLM systems on historical Devanagari printed documents.
- Legacy Typeface Error Pattern Analysis: Model predictions can be compared with verified ground-truth transcriptions to analyze character- and ligature-level recognition errors and identify recurring error patterns in historical Devanagari typefaces.