baligh-ged-crf-morph

Baligh Ged Crf Surface Morph V2

This is Baligh's morphology-aware Arabic grammatical error detection CRF. It uses Baligh preprocessing tokens plus the disambiguated morphological features from preprocessing output, and emits the mapped labels UC, OT, MO, SY, PC, MG, SP, and UNK.

Selected development operating point

• Training datasets: qalb14, qalb15
• Development datasets: qalb14, qalb15
• Threshold: 0.35
• Binary error F1: 0.778608
• Precision: 0.861525
• Recall: 0.710251
• False positives per 1,000 tokens: 25.620425

The threshold applies to the strongest non-UC marginal. The selected error category is emitted only when that marginal clears the threshold.

Files

• model.joblib: trusted sklearn-crfsuite model artifact
• manifest.json: model, feature, threshold, data, and dependency metadata
• label_mapping.json: QALB-to-Baligh mapping
• threshold_sweep.json: development operating-point sweep
• smoke_test.json: dependency-free inference probe
• requirements.txt: exact loading dependencies
• SHA256SUMS: integrity checks

Safety and licensing

Joblib uses pickle internally. Load this model only from a trusted repository and verify SHA256SUMS first. Keep the repository private until QALB-derived model redistribution and the intended model license have been reviewed.