GLM-OCR-uz

GLM-OCR (0.9B) fine-tuned for printed Uzbek book pages: Uzbek Cyrillic, Uzbek Latin and Russian.

It reads the text exactly as printed, including old typefaces that leave out the diacritics in ў қ ғ ҳ. It does not "modernise" the spelling.

Accuracy

The test set has 41 pages from 29 books that were not used in training: 18 Uzbek Cyrillic, 15 Uzbek Latin and 8 Russian. Pages are scanned at about 150 dpi.

The reference text was proofread by hand, starting from Gemini's output.

Text is normalised before scoring:

• markdown removed;
• hyphenated line breaks joined;
• apostrophes and dashes unified.

WER ↓	all	Uzbek Cyrillic	Uzbek Latin	Russian	size
GLM-OCR-uz (this model) + postfix.py	4.5%	5.7%	1.4%	7.4%	0.9B
uzocr-8b	3.9%	5.1%	1.8%	4.8%	8B
Gemini (API)	2.2%*	3.5%*	1.5%*	0.3%*	—

* The reference was drafted from Gemini's text, so Gemini's numbers may be slightly optimistic.

Uzbek Latin is better than both uzocr-8b and Gemini. Uzbek Cyrillic is close to uzocr-8b at about 1/9 of its size. Russian is weaker; it is not the target language.

Speed

Both models ran on the same 41 pages on one RTX 5090 with vLLM 0.28.

• This model: bf16, pages upscaled to a 1600 px long side.
• uzocr-8b: FP8, pages upscaled to a 2000 px long side (its best setting).

	GLM-OCR-uz	uzocr-8b	speed-up
latency, 1 page at a time	1.5 s/page	6.6 s/page	4.4×
throughput, 16 parallel requests	3.3 pages/s (≈11,800 pages/hour)	0.68 pages/s	4.8×
output tokens/s (16 parallel)	~3,000	~600
model weights in VRAM	~2.2 GB (bf16)	~9 GB (FP8)

With plain transformers (bf16, batch 1, no vLLM), it averages 5.0 s/page, uses 2.5 GB peak VRAM, and generates about 180 tokens/s.

Usage

Upscale small scans so the long side is about 1600 px. This matters a lot: on a hard test set, WER dropped from 17% to 8% with upscaling.

import torch
from PIL import Image
from transformers import AutoProcessor, GlmOcrForConditionalGeneration
from postfix import fix_mixed          # optional: maps stray Latin homoglyphs inside Cyrillic words

M = "Saidakmal/GLM-OCR-uz"
proc = AutoProcessor.from_pretrained(M)
model = GlmOcrForConditionalGeneration.from_pretrained(M, dtype=torch.bfloat16, device_map="cuda").eval()

im = Image.open("page.jpg").convert("RGB")
s = 1600 / max(im.size)
im = im.resize((int(im.width * s), int(im.height * s)), Image.BICUBIC)

msgs = [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": "Text Recognition:"}]}]
prompt = proc.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
x = proc(text=[prompt], images=[im], return_tensors="pt").to("cuda")
x.pop("token_type_ids", None)
out = model.generate(**x, max_new_tokens=3000, do_sample=False, repetition_penalty=1.05)
text = proc.decode(out[0][x["input_ids"].shape[1]:], skip_special_tokens=True).replace("<think></think>", "").strip()
print(fix_mixed(text))

vLLM:

vllm serve Saidakmal/GLM-OCR-uz --max-model-len 8192 --limit-mm-per-prompt '{"image":1}'
# chat/completions: image + "Text Recognition:", temperature 0, repetition_penalty 1.05

Training

The model is a full fine-tune of zai-org/GLM-OCR, trained in three rounds with 8-bit AdamW, batch 1 and gradient accumulation 8.

Real pages (about 26.6K) are scans of Uzbek and Russian books. They are labelled as follows:

• Uzbek Cyrillic: a word-level majority vote of three OCR systems (Gemini, uzocr-8b and an earlier version of this model). Its WER against the hand-checked reference is about 1.7%. 143 pages, where the systems disagreed, were checked by hand.
• Uzbek Latin and Russian: Gemini.

Pages in other Turkic Cyrillic orthographies are filtered out: Tajik, Karakalpak and others.

Synthetic pages: about 70K pages of rendered Uzbek and Russian text. They use many fonts, layouts and degradations, with exact labels.

Image size: pages are trained at a 1600 px long side, randomly scaled by 0.85–1.0.

Limitations

• The model is trained for printed text. Handwriting, tables and complex layouts are not specifically trained.
• Russian is noticeably weaker than Uzbek.
• The model sometimes misreads individual letters on very poor scans. That is where the larger uzocr-8b is still ahead.