SLAYER-OCR model registry
Model artifacts for reproducible SLAYER-OCR experiments. Every payload is accompanied by hashes and provenance. The public SLAYER Vision ONNX artifact remains at PiotrSty/slayer-vision-onnx.
- Source repository:
PiotrStyla/OCR_engine - Source commit:
6c246e0cd79d2a36443cd147e56964dd826a4293 - Registry generated:
2026-09-30T06:36:23.096415+00:00 - Visibility: public
- Integrity: see
MANIFEST.jsonlandMANIFEST.sha256
Historical Polish spelling and transcription variants are preserved. No normalization from historical glyphs or diacritics to modern spelling is implied.
Frozen Qwen historical adapter 0.5 — v1
SLAYER OCR Qwen3-VL hist 0.5 frozen — original unmerged LoRA weights, pinned base and inference recipe, exact recorded runtime and audited evidence.
Release: qwen3vl-hist-0p5-v1-20261013; registry receipt. Amended provisional fullpage DEV: 13.87% CER, 15 pages, no errors. Frozen test A: 18.79% CER, 36 pages including one failed/empty generation. Experimental reference only; no production promotion, SOTA or below-16% A claim.
This registry remains code-free. Runnable frozen code and original adapter weights are in the dedicated model snapshot above. Existing model artifacts remain unchanged.
Preserved Polish print engine — v1, 2026-10-16
SLAYER OCR print engine, frozen full snapshot:
OpenCV + TrOCR mixed-v3, full unchanged weights, tokenizer/processor, runnable
engine code and exact configuration. Tag: print-engine-v1-20261016;
registry receipt.
Współczesny druk PL: potwierdzono wykonanie pełnego pipeline'u na jednej syntetycznej stronie, nie jakość na reprezentatywnym zbiorze rzeczywistych współczesnych dokumentów. CPU smoke: 5/5 lines, 6 edits / 168 characters, CER 3.5714%, WER 24.00%, no LLM correction. The older 75-line 5.3307% CER control is historical ordinary print, not contemporary fullpage evidence. No production-readiness or SOTA claim; no existing model was overwritten.
This is separate from Qwen/EZH historical experiments. This registry entry contains only metadata; all runnable code and weights are preserved in the dedicated snapshot, with raw inputs/outputs, limitations and checksum proof.