blizzarman/polyglot-tutor-cefr-onnx

Polyglot Tutor — CEFR text classifier (ONNX int8)

XLM-R-base fine-tuned on UniversalCEFR English reference subsets (cambridge_exams, elg_cefr, cefr_sp, readme), exported to ONNX with dynamic int8 quantization for CPU inference. Document-level evaluation (chunk + expected-rank aggregation): macro-F1 0.798, QWK 0.936, adjacent accuracy 0.990 on 99 held-out documents. Full methodology, audits and caveats: https://github.com/arthur-diaz/polyglot-tutor (docs/evals/m1_classifier_report.md).

License is inherited from the training corpora (CC BY-NC-SA 4.0, non-commercial). Please cite the UniversalCEFR collection and the original corpus papers if you use this model.