Balinese Lontar Line-Level Transcription Corpus

A line-level image-text corpus for handwritten text recognition of Balinesescript on palm-leaf (lontar) manuscripts. It contains 144,706 line cropspaired with their transcriptions, drawn from 426 manuscripts. No line was manually annotated for this release. The corpus was built byre-aligning transcriptions that already existed on the open platformpalmleaf.org with their source images, using saturation-based leaf detectionand transcription-guided line segmentation. The split is frozen and assigned at the level of same-named sibling manuscriptgroups rather than individual manuscripts, because sibling manuscripts weremeasured to share text 13.4 times above chance. Users who reshuffle the splitwill leak text between training and test. Inter-transcriber variation on this corpus is 4.53% CER, measured from 310line pairs that were independently transcribed twice. Source material is from palmleaf.org and is licensed CC BY-SA 4.0. Thisderived dataset carries the same license. Ethical notes: colophons in many manuscripts contain the names of owners,scribes, and their home banjar. These names are part of the source text andcannot be removed without damaging the manuscript. Units containing modrescript, regarded as sacred, were deliberately excluded.