Repository navigation
DocLayout v3.0.0
DocLayout 3.0.0 adds optional local PP-DocLayoutV3 guidance to the existing GPT-6 Sol transcription pipeline. Sol still reads the full page and writes HTML. Confident, policy-qualified matches can use V3 rectangles and reading order; unmatched or ambiguous Sol content is retained. If layout preparation or inference fails, conversion continues with Sol and records the fallback.
Highlights
- Revision-pinned, hash-checked ONNX weights downloaded to an ignored local cache when absent; the layout extra requires Python 3.11+.
- GUI, CLI, API, PDF, and OCR conversion paths share layout guidance and diagnostics. Auto device mode can fall back to CPU; explicit CUDA does not silently switch devices.
- JSON, chunks, Markdown, OCR JSON, and annotations consume the resulting document. Annotations draw rectangles, not segmentation masks.
- Updated installation, configuration, architecture, OpenWiki, and interactive diagrams.
Install
The wheel and source archive are attached. PyPI publication remains deferred. For the base CLI without Git:
uv pip install "https://github.com/pypi-ahmad/DocLayout/releases/download/v3.0.0/doclayout-3.0.0-py3-none-any.whl"
To use V3 guidance, install the downloaded wheel with the layout extra, optionally adding gui or server. Configure Sol credentials before conversion.
Verification and limits
- 514 offline tests passed; 3 skipped. The uv lock check, Ruff correctness rules, and whitespace check passed.
- Both distributions built. The source archive excludes local cache content and includes the license, notice, and four prompt resources.
- An isolated Windows Python 3.13 wheel install with gui, server, and layout extras passed dependency and import checks.
- No live Sol, CPU-layout, or GPU-layout conversion was run for this release. Model accuracy and fresh-machine GPU behavior remain unverified. Repository-wide style Ruff and ty are not clean.