OpenThai 2.0 OCR Markdown

What it does​

• Keeps tables as tablesTables come out as HTML with rowspan and colspan, so the merged header cells of Thai government forms survive. 95.2 on 65 real Thai government and university pages, against 91.3 for its base model.
• Writes MarkdownHeadings, lists and paragraphs come out as Markdown, ready for a search index, a retrieval pipeline or a language model's prompt.
• Marks where each block isEvery block carries a bounding box and a class such as title, text, table or picture, in reading order.
• Reads columns in orderTwo- and three-column pages are read column by column instead of straight across.
• Runs on one GPU1.6 seconds a page on one NVIDIA H100 with vLLM, one page at a time. 0.63 billion parameters.

Measured​

In short: significantly better than Typhoon OCR V2 2B on Thai document tables and on new Royal Gazette pages, level with it on ThaiOCRBench tables and on printed text, and about 10 times faster at a third of the size. On 339 new Royal Gazette pages, published after our training data was collected, it read 98.7% of characters correctly, against 94.6% for Typhoon OCR V2 2B. The benchmarks, the scoring scripts and how the checkpoint was chosen are in the model card.

Run it​

hf download iapp/OpenThai2.0-OCR-Markdown --local-dir OpenThai2.0-OCR-Markdown
pip install vllm==0.26.0
pip install -e ./OpenThai2.0-OCR-Markdown
vllm serve ./OpenThai2.0-OCR-Markdown --served-model-name nemotron-parse --trust-remote-code

from nemotron_parse.inference import DocumentParser, GenerationError

with DocumentParser(base_url="http://localhost:8000/v1", model="nemotron-parse") as parser:
    try:
        doc = parser.parse("page.png", mode="single")
        print(doc["markdown"])  # Markdown; tables as HTML with rowspan and colspan
    except GenerationError:
        pass  # raised if a page is cut off at the token limit; see Limits

Tested with vLLM 0.26.0, transformers 5.14.1, PyTorch 2.11 and Python 3.10. It runs anywhere Nemotron 3.5 Parse runs.

Prefer a hosted service? The OpenThai 2.0 OCR Markdown API runs this model for you, in Thailand, at 0.15 IC a page. For plain text, Word, PowerPoint, Excel or a searchable PDF, OpenThai 2.0 OCR Text starts at 0.049 IC a page.

Limits​

• Repetition loops. None in our tests: no page was cut off by a loop among 2,270 printed-text pages, 193 ThaiOCRBench table images, 65 Thai document table pages and 6,137 development pages. If another page is cut off at the token limit, DocumentParser raises GenerationError, so the page can go to another reader.
• Photographed single words and very short crops read worse than full pages: 96.4 on the OCR104 photographed document crops, against 99.8 for Typhoon OCR V2 2B.
• Dot-matrix print (bank passbooks, impact-printed forms) reads less well than ordinary print: about 89% character accuracy on our dot-matrix development crops; the crops are generated and there is no real dot-matrix test set yet, so expect less on real passbooks.
• Handwriting is read at the level of single text lines; full handwritten pages are not a target.
• Scene text and photographs of screens are not targets.
• Charts are read as the base model reads them; they were not part of the fine-tuning.

Training​

The text labels are exact, never another model's reading: 54,596 training rows, mostly synthetic Thai documents (forms, memos, minutes, schedules and merged-cell tables, rendered with scan and phone-photo damage, plus two- and three-column pages) and born-digital Thai government pages, the Thai parliament's budget documents and the Royal Gazette, whose text layer gives the exact text, plus tables of other kinds: public financial and scientific tables (FinTabNet and PubTabNet), and synthetic Thai receipts, statements and lab reports, some with every digit in Thai. They include 420 pages on which earlier fine-tunes ran into a repetition loop, with the loop as the thing not to write. Trained on one H100 for one epoch of 1,707 steps; the release is the last checkpoint, used as trained, with none of the base model's weights blended in. It was picked by a rule fixed before any candidate was scored, on development sets only. No benchmark page was trained on. The mix, the recipe and the rule are in the model card.

Licence​

OpenMDW License Agreement, version 1.1, the licence of its base model. OpenThai 2.0 OCR Markdown is a Thai fine-tune of NVIDIA Nemotron 3.5 Parse (0.63B); NVIDIA's licence, notices and original model card are included unchanged in the repository. The tokenizer is MIT-licensed.

Built on NVIDIA Nemotron.

Built by iApp Technology. Trained and evaluated on an NVIDIA H100 GPU provided by Siam AI Corporation.