Commit History
18 days ago
-
Release Sanskrit OCR pipeline v1.2
BLLA v1.2 continues from v1.0 with slant/curve augmentation (rotation up to 6 degrees, row-bending warps applied to page and target together), a 12-pixel topline target, and slanted or curved training pages counted twice. The pipeline runs it at a 1,200-pixel input height and drops toplines shorter than 20% of the page's median topline. On the 229-page test split, exact line counts rise from 142 to 228 pages and pooled CER falls from 3.83% to 3.64% (M00638: 3.66% to 3.45%) in the same evaluation run. Settings were confirmed on 337 held-out pages not used to choose them. The recognizer is unchanged (PP-OCRv6 v1.0). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>bf37109
-
chore: streamline model repository
8e5390b