Q-Invoice-50M-Sovereign
Research checkpoint for invoice field extraction in the Qovaryx specialist family. The repository name is a historical identifier.
Status — October 3, 2026: Qovaryx is on the back burner. Desktop downloads are paused; there is no current public installer, purchase, or trial. These repositories remain as research references. Published weights are separate from desktop product availability.
Intended task
Extract structured fields from supplied invoice text. This describes the intended task, not a guarantee of correct, valid, or safe output.
Validate fields against the original invoice. OCR, PDF rendering, and accounting validation are separate components.
Published artifacts and dependencies
pytorch_model.pt: PyTorch checkpoint.- config.json: architecture and historical evaluation metadata.
tokenizer.json: tokenizer for this checkpoint.- release.json: historical release receipt.
The configuration describes a custom decoder with width 512, 12 layers, and vocabulary size 32,000. Use this checkpoint's configuration and tokenizer together. The similarly named public random-initialized scratch base is a separate artifact; its vocabulary and weights are not interchangeable with this release.
The historical loader imports FinanceDecoder and DecoderConfig from bleeding_edge.model.decoder. That compatible source package is not included in this repository. The compiled Q-Office-Suite runtime is not a substitute for a source-package installation. This is not a standalone Transformers AutoModel release.
Evaluation scope
Previous cards presented perfect scores from small internal task fixtures as broad accuracy claims. That framing has been removed. The configuration and release receipt retain the original metadata; this documentation update does not independently rerun or certify those evaluations. Some earlier prose and metadata used different sample counts.
Those fixtures do not establish performance on arbitrary inputs, unfamiliar documents, different schemas, or deployed workflows. Valid JSON and correct answers require separate validation. There is no current production-readiness or general-accuracy claim.
License
Apache-2.0, as recorded in the existing model metadata. This statement does not certify the provenance or license of every historical training example, separate runtime component, or downstream dependency.
Project and support
J.E. Herizon LLC, doing business as JE Horizon.