tre-ocr: round 1 recogniser

A PaddleOCR PP-OCRv6 medium text recogniser, fine-tuned for Vietnamese, from ternlabs/tre-ocr: OCR for Vietnamese documents, with the accuracy measured, not assumed.

What it is for

Reading scanned Vietnamese stock-exchange disclosures (resolutions, notices, explanations filed by HOSE and HNX companies). It is meant to run inside the tre-ocr tool, which pairs it with PaddleOCR's stock recogniser on the same text boxes: where their digits differ the stock line is used, and a number is reported only when both agree with confidence of at least 0.95. Otherwise the number is left empty ([?]), never guessed.

Used alone, outside that pairing, it has not been measured.

Results (set F, measured once, pass marks fixed in advance)

40 scanned pages, none seen in training, at most 2 per company.

Mark Needed Result 95% interval
Number recall โ‰ฅ 98% 98.7% (607 of 615) 97.5 to 99.3
Wrong when sure โ‰ค 1% 0.19% (1 of 535) 0.0 to 1.1
Coverage โ‰ฅ 85% 87.4%
Company field โ‰ฅ 90% 91.7% (22 of 24) 74.2 to 97.7

The candidate passed each mark, but the intervals reach past them: close to the marks, not clear of them. Tone marks were scored against a reading no person checked, so they are reported in METHOD.md and never used as a pass. Full method, sets and limits: METHOD.md.

Limits

  • Financial statements have not passed; the tool refuses them.
  • Handwriting, other document kinds and other scan conditions are not measured.

Training

Fine-tuned only on typed public filings rendered to look scanned. Every test-set file, and every page naming a test-set company, was left out.

Use

git clone https://github.com/ternlabs/tre-ocr && cd tre-ocr
pip install -r requirements-reader.txt     # needs poppler-utils
bash tools/get_weights.sh                  # fetches and hash-checks these weights
python -I tools/read_pdf.py some.pdf --out cache/drafts

Files here: round1-rec.tar.gz, SHA-256 010c5757be568a258d6657402a3acbbf6450f46dd978840e4f5a298e1e22fd5e, the same archive get_weights.sh checks.

Licence

Apache License 2.0. Published by Tern Labs.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for ternlabs/tre-ocr

Finetuned
(4)
this model