uzocr-8b: OCR for scanned Uzbek documents

uzocr-8b reads scanned pages of Uzbek books, dissertations and dissertation abstracts and returns their text as Markdown.

• It handles both Uzbek scripts, Cyrillic (including ў, қ, ғ, ҳ) and Latin (oʻ, gʻ, ʼ), on typewritten and printed pages, including low-quality scans.
• It preserves the page structure: headings, paragraphs, tables, footnotes, formulas (LaTeX) and the table of contents.
• The model runs fully locally on a single GPU.

O'zbekcha. uzocr-8b — skanerlangan o'zbek kitoblari, dissertatsiya va avtoreferatlar betini o'qib, matnini Markdown ko'rinishida qaytaradigan model. U kirill va lotin yozuvidagi, mashinkada yoki bosmaxonada terilgan, sifati past skanlar bilan ham ishlaydi. Sarlavha, jadval, snoska, formula va mundarijani saqlaydi. Bitta GPU'da to'liq lokal ishlaydi.

Results

Evaluation set: 300 pages sampled from 11 held-out Uzbek documents (dissertations, dissertation abstracts, a textbook, a monograph; Cyrillic and Latin; printed 1999–2023; one low-quality scan), none of which were used for training or tuning. Scores are on the 284 pages whose reference covers the full body text (58,968 words). Normalised word and character error rates (lower is better; case and punctuation ignored); all systems were given the same page images and, for the VLMs, the same prompt; single run each.

95% bootstrap interval for uzocr-8b WER: [1.04, 1.86] over pages, [0.98, 1.91] over documents. The paired difference to the base model is 10.6 points (interval [8.7, 12.8]). On the Cyrillic pages, 80% of the base model's single-letter errors replace an Uzbek letter by its Russian look-alike (қ→к, ҳ→х, ғ→г, ў→у); after fine-tuning these confusions are rare and symmetric. Details, the evaluation protocol and a training-stage ablation are in the paper below.

Speed: 1.15–1.7 s per page with vLLM (bf16, 16–24 concurrent requests) on one RTX 5090; about 32 s per page with llama.cpp Q8_0 on a 12 GB RTX 3060.

Usage

The model was trained with the system prompt below. Use it verbatim, with greedy decoding (temperature=0).

You are a high-precision OCR engine for scanned BOOK pages written in Uzbek (Latin or Cyrillic), sometimes mixed with Russian or English. Pages may contain printed or typewritten text, handwritten notes, tables, poems, footnotes, stamps and images.

TASK: Transcribe the page image EXACTLY as written and return clean Markdown.

RULES:
1. Verbatim only. Do NOT translate, summarize, paraphrase, modernize spelling, fix grammar or complete sentences. Cyrillic stays Cyrillic, Latin stays Latin. Keep Uzbek Cyrillic letters ў қ ғ ҳ exactly as printed.
2. Uzbek Latin: write oʻ and gʻ with U+02BB (ʻ) and the tutuq belgisi with U+02BC (ʼ). Never ó, ğ or backticks.
3. Handwriting: if it is a margin or interlinear note added to printed text, write it as {qoʻlyozma: ...} where it appears.
4. Uncertain word → best reading + [?] (Toshkent[?]). Unreadable → [oʻqib boʻlmaydi]. Never guess silently.
5. Book structure:
   - Chapter titles → "# ", section titles → "## ". Only for real visual headings.
   - Join all lines of a paragraph into one line; separate paragraphs with a blank line. A word split by a hyphen at a line end is joined (kitob-/lar → kitoblar); real hyphenated words keep the hyphen.
   - Poems: keep every line break; separate stanzas with a blank line.
   - Footnotes at the bottom → write at the end as [^n]: text, and keep the marker [^n] in the body where the number appears.
   - Tables → Markdown tables with all rows and columns.
   - Table of contents: write "title — page"; never reproduce dot leaders.
   - Drop caps (large first letter) → merge into the word.
   - Images/illustrations → [rasm: short caption if printed].
6. Only if the page really has a running header/footer or page number, put it on a separate first/last line prefixed with "HEADER:" / "FOOTER:". If there is none, do not write these lines.
7. Output ONLY the transcription. No comments, no code fences.

User message: the page image plus one of these texts:

• Cyrillic: Transcribe this page. This book is printed in Uzbek CYRILLIC script: keep every word in Cyrillic exactly as printed; never transliterate to Latin.
• Latin: Transcribe this page. This book is printed in Uzbek LATIN script: keep every word in Latin exactly as printed; never transliterate to Cyrillic.

Render pages at about 200 dpi (up to about 2.6–2.9 megapixels).

vLLM (recommended)

vllm serve JahongirB/uzocr-8b --served-model-name uzocr-8b --dtype bfloat16 --max-model-len 8192 \
  --max-num-seqs 24 --limit-mm-per-prompt.image 1 \
  --mm-processor-kwargs.min_pixels 200704 --mm-processor-kwargs.max_pixels 2600000 \
  --generation-config vllm

import base64, openai

SYSTEM_PROMPT = "..."  # the system prompt above
client = openai.OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="-")
img = base64.b64encode(open("page.png", "rb").read()).decode()
r = client.chat.completions.create(
    model="uzocr-8b", temperature=0, max_tokens=4096,
    messages=[{"role": "system", "content": SYSTEM_PROMPT},
              {"role": "user", "content": [
                  {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{img}"}},
                  {"type": "text", "text": "Transcribe this page. This book is printed in Uzbek CYRILLIC script: keep every word in Cyrillic exactly as printed; never transliterate to Latin."}]}])
print(r.choices[0].message.content)

llama.cpp (12 GB GPU)

Use uzocr-8b-GGUF:

llama-server -m uzocr-8b-v3b-Q8_0.gguf --mmproj mmproj-uzocr-8b-v3-F16.gguf -c 8192 -np 4 --temp 0

Requirements

• vLLM, bf16: about 17 GB for the weights plus the KV cache. With 24 concurrent requests, peak VRAM was 30 GB on an RTX 5090.
• llama.cpp, Q8_0: 8.7 GB for the model plus 1.2 GB for the vision projector.

Training

• Data: page images of Uzbek documents (Cyrillic and Latin, typewritten and printed, scans of varying quality) paired with Markdown transcriptions in the format of the system prompt above. Documents used for evaluation were excluded from training; this was checked with text-shingle overlap.
• Method: LoRA, r = 32, alpha = 64, dropout 0.05, on all attention and MLP projections of the language model. The vision encoder was frozen.
• Optimization: loss on answer tokens only, cosine schedule, gradient accumulation 8. Three stages (lr 1e-4, then 5e-5 and 5e-5). This release is the last checkpoint of stage 3.
• Compute: about 17.6 GPU-hours on a single NVIDIA RTX 5090.

Limitations

• Remaining errors are mostly single-letter confusions on faint or broken print, including the Uzbek-specific letters ў/қ/ғ/ҳ (e.g. қ ↔ к, ҳ ↔ х) and the fusion or splitting of words. The model reproduces the source as printed, including the source's own misspellings.
• Pseudo-graphic typewriter tables and some formula pages can cause repetition loops (the released checkpoint looped on none of the 284 evaluation pages under greedy decoding, but earlier checkpoints did). A loop check on the output (repeated tail or repeated 8-grams) and a retry with a small repetition penalty are recommended; the paper gives the exact criterion.
• Pages that mix scripts (e.g. an English summary page in a Cyrillic book) may be transliterated into the main script.
• Handwritten text is not a target of the model.
• Russian pages are readable, but the model was not tuned for them.

Paper and citation

Boltayev, J. (2026). Models and Algorithms for Reliable Recognition of Low-Quality Scanned Uzbek Texts Based on a Locally Adapted Multimodal Model. Preprint (arXiv identifier to be added); submitted to a journal.

@misc{boltayev2026uzocr,
  author = {Boltayev, Jahongir},
  title  = {Models and Algorithms for Reliable Recognition of Low-Quality Scanned Uzbek Texts Based on a Locally Adapted Multimodal Model},
  year   = {2026},
  note   = {Model: Hugging Face, doi:10.57967/hf/10800; GGUF: doi:10.57967/hf/10801}
}

Model DOIs: 10.57967/hf/10800 (this repository), 10.57967/hf/10801 (GGUF).

License

Apache-2.0, the same as the base model.