lightonocr-3-0.8b
Source-pinned OCR model data mirrored by Nile Academy for Wekala consumers. Original development and model rights remain with the publishers listed below.
Complete selected declarative source checkpoints and original model cards are preserved under original/<source-commit>/. Root files, or det/ and rec/, provide the selected deployment data. Original Paddle inference checkpoints are retained separately from official ONNX exports. No Python source, pickle checkpoints, inference servers or executable model code is included.
Qualification
Publication verifies immutable file identities, not OCR accuracy, native ports, quantization fidelity, device fit or Desktop/mobile/browser acceptance. The models are not bundled with the Wekala core. Consumers supply approved assets to qualified compiled adapters. Arabic support is preferred but never a model eligibility gate. Published language metadata does not establish tested quality in any language. No cloud fallback is selected.
Rights and provenance
Upstream model cards declare apache-2.0. Their exact bytes and any supplied notices are preserved with the original checkpoint. LICENSE retains an upstream license file when supplied; otherwise it contains the standard grant text corresponding to the publisher declaration. No copyright holder is invented. See publication.json for source commits and file identities. Engine code and any converted model assets need separate records.