Issues

• Upright Odia scans are turned 180° by orient.onnx (18 of 30 Wikisource pages), so the output comes out in other scripts

  datalab-to/indra#32 · Filter by author CyberNativeAI opened on Oct 28, 2026

• 60° rotated table page returns empty OCR text after incorrect orientation/deskew

  datalab-to/indra#29 · Filter by author lihongjie0209 opened on Oct 26, 2026

• UnicodeDecodeError loading math models on Windows (read_text without encoding)

  datalab-to/indra#27 · Filter by author ArpitKadam opened on Oct 25, 2026

• Pointed Hebrew texts recognition

  datalab-to/indra#26 · Filter by author mkpelletier opened on Oct 23, 2026

• Script classifier crashes on a line crop narrower than 32 px (input_shape {1,0}: ReduceSum on CUDA, Gemm on CPU with onnxruntime 1.19.2)

  datalab-to/indra#24 · Filter by author wuxuedaifu opened on Oct 23, 2026

• CUDA: GPU memory grows with every page and is never released (ONNX Runtime arena keeps its largest allocation)

  datalab-to/indra#22 · Filter by author wuxuedaifu opened on Oct 23, 2026

• --math fails forever when the model cache was primed without math (partial HF snapshot is trusted)

  datalab-to/indra#18 · Filter by author DylanMing opened on Oct 22, 2026
  1

• Pass known page rotation through convert(), and make ORIENT_MIN configurable

  datalab-to/indra#17 · Filter by author Amixx opened on Oct 20, 2026

• Spanish Ñ is lost with the Impact font (NIÑO → MINO, SEÑOR → SENOR); same page in Arial Bold reads 8/8

  datalab-to/indra#16 · Filter by author EliasMatDev opened on Oct 20, 2026

• Model-side speedups left after the native kernels (profiling evidence)

  datalab-to/indra#15 · Filter by author karanvora2599 opened on Oct 20, 2026
  1

• Digital PDFs: a page's first read in a process can differ (pdfium character boxes change after a render)

  datalab-to/indra#14 · Filter by author karanvora2599 opened on Oct 20, 2026

• Word-level bounding boxes

  datalab-to/indra#7 · Filter by author mkrzywda opened on Oct 17, 2026