| # Printed development control: 2026-09-21 |
|
|
| Source: user-provided `printed-dev-20260921T103012950790Z.zip`. |
| All 23 checksums verified; no unlisted payload files. Selection reproduced from |
| archived validation/test metadata with the original selection and overlap guards. |
| Metrics independently recomputed from aligned prediction IDs and references. |
|
|
| ## Results |
|
|
| 15 regions from three development pages/collections; 474 normalized reference |
| characters. Both models returned 15 nonempty outputs, zero recorded errors and |
| zero flagged truncations. Tesla T4, Torch 2.10.0+cu128, Transformers 4.57.6. |
|
|
| | Model | CER | WER | Lowercase CER (diagnostic only) | |
| |---|---:|---:|---:| |
| | microsoft/trocr-base-printed | 69.4093% | 89.5522% | 25.9494% | |
| | PiotrSty/trocr-pl-mixed-v3 | 23.4177% | 70.1493% | 20.2532% | |
|
|
| Primary metrics use NFC + whitespace only. The additional diagnostic applies |
| Python `lower()` after normalization, not casefold or historical spelling repair. |
| Microsoft frequently emits uppercase text. Thus much of the 45.99 percentage-point |
| primary CER gap is case error; the lowercase gap is only 5.70 points. Retain both |
| raw predictions and primary metrics; do not describe this as a general threefold |
| recognition improvement. |
|
|
| | Collection | Regions | Base CER | Mixed-v3 CER | |
| |---|---:|---:|---:| |
| | Choragiew_FT | 5 | 72.11% | 32.65% | |
| | Relacja_koronacji_FT | 7 | 74.09% | 17.00% | |
| | Wiesc_FT | 3 | 50.00% | 26.25% | |
|
|
| ## Interpretation and limits |
|
|
| Mixed-v3 is a promising candidate for further development evaluation, not a |
| production-ready historical recognizer. Both models still misread historical |
| glyphs and decorative type; no model selection for the frozen benchmark follows |
| from this small test. Region scores cannot be compared directly with full-page |
| Kraken/Tesseract scores. |
|
|
| The references include three private-use characters and one replacement character |
| across three regions. They remain in scoring. References were not manually audited. |
| Only 15 heading/publication-detail regions, from three pages, were evaluated; |
| upstream model-training overlap and near-duplicates remain unaudited. |
|
|
| Visual spot-check: Choragiew r003 is a single historical blackletter line with |
| small adjacent-line fragments at the lower edge; Relacja r016 is one legible line. |
| This is not a visual audit of all 15 crops. In r016 mixed-v3 reads almost all the |
| text but predicts a dotted z in place of the accented a in the opening word. |
|
|
| Next: prepare a larger, manually checked development set of genuine body-text |
| lines, stratified by typeface/document, with audited references and document-level |
| separation. Freeze selection before evaluating further preprocessing or training. |
| Keep the 36-page frozen test untouched until development choices are finalized. |
|
|
| Raw files remain local in `data/printed-dev-20260921/`; `review.html` shows crops, |
| references and both outputs. No images or predictions were published by this audit. |
|
|
