Skip to content

Releases: PracticalSwan/OCR-Model

Release list

OCR Model v1.1.0 - domain-adapted OCR upgrade

OCR Model v1.1.0 - domain-adapted OCR upgrade Latest
Latest

Choose a tag to compare

View all tags
@PracticalSwan PracticalSwan released this 29 Jul 09:52
v1.1.0-ocr-upgrade
b7a2d10

This release publishes the corrected portable package from the domain-adapted OCR upgrade merged through PR #2, with the same stable v1.1.0-ocr-upgrade version and asset names.

Package provenance

  • Package build commit: 36ca41ee48dabd42bbb26f33cd490410a7f00c57
  • Clean Git candidate-tree SHA-256: 72ef8b001b05aa5ddabdb6218ee4a4236a515999638ba85675e2779f00d5d77a (522 candidates)
  • Selected LayoutXLM checkpoint SHA-256: f257538849bd2067a9df9df83385aa10ae468d0499510fb0621a03a5f0155180
  • Archive size: 1,159,079,957 bytes
  • Archive SHA-256: 4584bc6d9e782a9e50c5dc801ac3b6e1e633ef40a57f64763d5765ce285c546a
  • Archive inspection: 183 entries; CRC, single-root, duplicate-name, absolute-path, traversal, and 182-record payload-manifest checks passed
  • Privacy audit: no raw/private data, private outputs, credentials, secret patterns, or private-filename inventory matches

Executed release checks

  • Host suite: 417 passed, 3 skipped; Python compilation passed
  • Fresh Windows CPU setup and doctor passed with Paddle 3.3.0, PaddleOCR 3.7.0, Torch 2.8.0+cpu, and Transformers 4.57.6
  • Windows CPU public extraction passed; private extraction used an opaque run ID, ignored the public output override, and created no visualization
  • Gradio 6 public preview/gallery/archive serving passed; private preview/gallery/archive suppression and upload cleanup passed
  • Docker Desktop linux/amd64 build, loopback GUI probe, in-container doctor, and one-page CPU extraction passed; the exact test container, image, and network were removed afterward
  • Physical macOS hardware was not tested; macOS support remains the documented Docker Desktop CPU route

Model selection

The global OCR default remains the original PaddleOCR detector and general recognizer. The custom Thai recognizer is synthetic-only and available through explicit custom/adaptive experiments with explicit generic-layout fallback. The locked 1,760-page image-to-JSON evaluation completed without failures but missed the requested accuracy targets, so this remains an academic pre-model that requires human review.

Install

  1. Download OCR_Model.zip and OCR_Model.zip.sha256.
  2. Verify the checksum.
  3. Extract the ZIP to a writable folder.
  4. On Windows, run setup_windows.bat, then launch_windows.bat.
  5. On macOS, use the documented Docker Desktop Linux/AMD64 CPU route.

The historical v1.0.0-build-week release is intentionally preserved and has not been overwritten.

Assets 4
Loading

OCR Model - OpenAI Build Week portable release

OCR Model - OpenAI Build Week portable release

Choose a tag to compare

View all tags
@PracticalSwan PracticalSwan released this 20 Jul 15:52
v1.0.0-build-week
e47023d

OCR Model — OpenAI Build Week portable release

Portable, local-only OCR and document information extraction for Windows and
the Docker-backed macOS route. The archive includes the exact final LayoutXLM
and PaddleOCR weights, one-step CLI and GUI launchers, diagnostics, a synthetic
demo sample, manifests, setup instructions, the MIT license for original
code/documentation, and the solo-maintainer contribution policy. It does not
use an OpenAI API key.

This release repairs the local GUI with:

  • an image preview and first-page PDF preview;
  • exactly one compact extraction progress indicator;
  • independently scrollable OCR-text and run-log panes;
  • stale-result clearing when a new document is uploaded; and
  • ANSI-free run logs.

Verified before publication:

  • clean 1,152,835,265-byte archive with one OCR_Model/ root and no raw or
    private data;
  • archive SHA-256
    c6c874f5b0879478497c9a33529f6416d48be60d586197fb625540d795f9ec6b;
  • source commit
    e47023de2a201092df6fd3393ec297b2835e0a50;
  • ZIP CRC, single-root, required-file, and privacy audits;
  • portable doctor with OCR, LayoutXLM, and K-Means import probes;
  • full CPU extraction of the bundled sample with the expected five fields;
  • native Windows GPU doctor, CLI extraction, and packaged-GUI extraction;
  • rebuilt Docker Linux/AMD64 CPU doctor and full extraction with exact field
    values/methods, OCR text, entity/relation triplets, and rotation parity;
  • host test suite: 244 passed and 2 environment-dependent skips; and
  • final LayoutXLM checkpoint SHA-256
    34c7a26e78d6285a2739e1b61839eadfd0e686ccbcf57f9cb47997c12cef2189.

Physical Apple hardware was not available. macOS support is the Docker Desktop
linux/amd64 CPU path and may be slow on Apple Silicon.

Loading