OCR experiment registry
Collection
Versioned registry of SLAYER-OCR datasets, models and experiment evidence. Code remains on GitHub. • 4 items • Updated
Dataset Viewer
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): ('json', {}), NamedSplit('validation'): ('imagefolder', {}), NamedSplit('test'): ('json', {})}
Error code: FileFormatMismatchBetweenSplitsError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
SLAYER-OCR experiment evidence registry
Evidence, metrics, predictions, environment records, result reports and the non-code artifact timeline recovered from GitHub. Failed and negative experiments are intentionally retained.
- 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.
Qwen3-VL historical adapter strength — DEV checkpoint
Audited result, predictions and inference recipe: DEV-selected strength 0.5; region DEV CER 13.20%, provisional fullpage DEV CER 19.90%. No new training, no frozen test A measurement and no model promotion. The below-16% frozen fullpage test A target is not established.
Frozen historical Qwen adapter 0.5 — v1, 2026-10-13
Frozen model and recorded runtime; code-free evidence snapshot.
Amended provisional fullpage DEV v4: 13.87% CER, 15 pages, no errors. The unchanged frozen 36-page test A is 18.79% CER, including one failed/empty generation. This is a frozen experimental reference, not promotion or a below-16% claim. Original measurements, original checkpoint bytes, scope limits and negative Paddle control are preserved. CPU scores and packaging were verified; no new OCR inference or training was run to freeze the release.
The previous DEV checkpoint above retains its original provisional reference and scores; it is not overwritten by the amended v4 rescore. This evidence registry remains code-free; executable exact runtime files live only in the dedicated model snapshot.
Historical glyph fidelity prompt — measured diagnostic DEV candidate
Predictions, independently recomputed scores, provenance and offline-review diagnostics. All fifteen pages completed on Tesla T4 with no missing or empty outputs. Candidate CER is 13.644758283541553%, versus 13.872895165670832% for saved frozen-v1 outputs: 21 fewer character errors across 9,205 reference characters. Five pages improve, seven regress and three tie on CER. References remain a single-review, non-adjudicated, scope-incomplete draft; the baseline was not rerun in the same session. This is a candidate for review, not a new frozen release, prompt-only causal proof, test A result or promotion. The code-free publication preserves raw predictions and the original local archive identity; the original ZIP contains executable sources and is not uploaded to this registry. Frozen model weights and release pointers are unchanged.
- Downloads last month
- 992