harsha-desaraju/telugu-line-ocr-bench
Telugu Wikisource OCR — human-verified line crops 1044 single-line crops from Telugu Wikisource page scans, each with a transcription checked against the image by a human. Grayscale, height 64px, width a multiple of 8 — the form the encoder consumes. Columns column meaning image the line crop text gold transcription, human-verified n_graphemes akshara count of text (regex.\X) has_english text contains a Latin-script letter. Digits/punctuation do… See the full description on the dataset page: https://huggingface.co/datasets/harsha-desaraju/telugu-line-ocr-bench.
Telugu Wikisource OCR — human-verified line crops
1044 single-line crops from Telugu Wikisource page scans, each with a transcription checked against the image by a human. Grayscale, height 64px, width a multiple of 8 — the form the encoder consumes.
Columns
Script content
Every line is free of Latin letters — the set is Telugu plus numerals and punctuation. has_english is retained so the filter is auditable, and is False for every row when the dataset is built with DROP_LATIN.
Intended use
Evaluating Telugu line-level OCR. The labels carry no dependence on any particular recogniser, so engines can be compared fairly against them.
Source
Built from Telugu Wikisource proofread pages, segmented into lines and aligned, then hand-corrected. Text inherits Wikisource's CC BY-SA 4.0.