Reproducibility package for "Models and Algorithms for Reliable Recognition of Scanned Uzbek Cyrillic and Latin Texts by Staged Low-Rank Adaptation of a Locally Deployed Multimodal Model"
Annotatsiya
Reproducibility package for the article "Models and Algorithms for Reliable Recognition of Scanned Uzbek Cyrillic and Latin Texts by Staged Low-Rank Adaptation of a Locally Deployed Multimodal Model" (submitted to Information, MDPI). The model is uzocr-8b (https://huggingface.co/JahongirB/uzocr-8b, DOI 10.57967/hf/10800). Contents: (1) zenodo_metrics_code_tables.zip: README; page-level results for the 300-page benchmark (document, page, coverage flag, reference word and character counts, normalised WER/CER and output word count for five systems: uzocr-8b model-only, Qwen3-VL-8B base, Qwen3.6-35B-A3B, Tesseract 5 at 200 dpi, uzocr-8b full pipeline); aggregate tables added in the revision (full pipeline vs single pass, Tesseract 5 at 300 dpi, uzocr-8b at 4.5 MP/300 dpi, 30-page check on unseen dissertations); evaluate.py with the normalisation, WER/CER and page-bootstrap interval exactly as used in the paper. (2) zenodo_sample_pages_e03.zip, zenodo_sample_pages_e04_part1.zip, zenodo_sample_pages_e04_part2.zip: 20 benchmark pages from two public dissertation abstracts (documents e03 and e04, 10 pages each) with the page image, the human-checked reference text and the raw output of each system. The remaining 280 benchmark pages are parts of copyrighted dissertations and books and are not distributed; only their page-level metrics are released.
Hali tarjima qilinmagan