grammar-ur v2 — Urdu grammar correction (GECToR, ONNX)

A token-tagging grammatical-error-correction model for Urdu (mdeberta-v3-base encoder, 3-round GECToR schedule), packaged as a single-file int8 ONNX tagger with its tokenizer artifacts — no Python runtime needed at inference.

Measured quality

Held-out gate: 1,500 corrupted Urdu sentences (six balanced single-error classes — homophone letter sets, Arabic↔Urdu letter confusions, keyboard adjacency, transpositions, space joins, real-word swaps), decoded from this exact artifact, scored by span overlap:

metric value
F0.5 0.640
Precision 0.678
Recall 0.524

License pre-release gate (Clause 16(e))

Per the Ribose⇄CLE Linguistic Resource License Agreement (executed 2026-10-29), this model passed extraction/memorization testing before release: zero novel ≥25-word corpus reproductions in decodes; membership inference AUC 0.485 (no signal); no completion/extraction surface; no WordNet text leakage in the released tokenizer artifacts. Test records are retained for 2 years.

Training data

Trained on a synthetic corruption pack derived from, and evaluated against held-out slices of:

CLE Urdu Digest Corpus 1M (International Package), Catalogue No. CLE12T003; CLE Urdu WordNet (International Package), Catalogue No. CLE12L001. Center for Language Engineering, Al-Khawarizmi Institute of Computer Science, University of Engineering and Technology, Lahore, Pakistan.

The Linguistic Resources themselves are proprietary to CLE-KICS, UET, are not redistributed by this repository, and are not extractable from the released weights (Clause 16(e) testing above). Third-party use of these weights, including in commercial products, requires no separate license from CLE-KICS, UET (agreement Clause 9).

Serving

Load through the kotoshu gem (Kotoshu.setup(:ur)); whitespace tokenization, Arabic→Urdu letter canonicalization (ك→ک, ي→ی), and the CRULP phonetic keyboard layout for proximity ranking are handled by the gem's Urdu language module. Decoding: 4 rounds, zero-confidence gate.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support