Darmm/darmm-ocr-kazakh-vl-3b
Image-Text-to-Text • 4B • Updated • 22
OCR for printed Kazakh (Cyrillic): models and synthetic training datasets
Note v2 — Qwen2.5-VL-3B fine-tune, reads full pages; CER 0.5–1.5%
Note v1 — 0.3B TrOCR-style, word/line crops; superseded by the VL-3B model
Note 282k crops+pages: clean, hard (scan-degraded), page tiers
Note original 200k clean crops (included in v2 as the clean tier)