Logo Lanfrica

Samielakkad/AI-NLP-Darija-Classifier

Domaine:

natural language processing

Type de record:

model
Créateur:
Sam
Hôte:
AI + NLP · production Darija (Moroccan Arabic) intent + city classifier, Hugging Face–ready --- license: apache-2.0 language: - ar - fr - en tags: - darija - moroccan-arabic - text-classification - intent-classification - low-resource - arabic-dialects - service-marketplace library_name: transformers base_model: xlm-roberta-base pipeline_tag: text-classification datasets: - samielakkad1/jakma-darija-intents metrics: - accuracy - f1 model-index: - name: jakma-darija-classifier results: - task: type: text-classification name: Intent + City Classification (Darija) metrics: - type: accuracy value: 0.92 name: 5-fold CV accuracy (initial baseline) - type: f1 value: 0.89 name: macro-F1 --- # AI + NLP · jakma-darija-classifier > Pass-1 intent + city classifier for jak.ma — a production Moroccan-Arabic (Darija) service marketplace. Classifies natural-language queries in Darija / Arabic / French / English / mixed-script into one of 12 trade categories and one of 15 Moroccan cities, with confidence-gated routing. **Author:** Sami El Akkad · Tsinghua SIGS, MSc Artificial Intelligence **Email:** sam25@mails.tsinghua.edu.cn **Production system:** jak.ma (live since Apr 2026 · ~3,400 daily queries · p50 1.18s · $0.0008/query) **Methodology:** pm-frameworks-darija · jak-ma-eval-suite --- ## What this model does Given a user query in any of: **Darija (Latin or Arabic script), Modern Standard Arabic, French, English, or code-switched mix**, this model predicts: 1. **Trade category** — one of 12 fixed classes (plumber, electrician, tiler, painter, carpenter, mason, mechanic, electronics, AC technician, gardener, cleaner, mover) 2. **City** — one of 15 Moroccan cities (Casablanca, Rabat, Salé, Tangier, Marrakesh, Agadir, Fes, Meknes, Oujda, Kenitra, Tétouan, Nador, Béni Mellal, El Jadida, Mohammedia) 3. **Confidence band** — `high` (route directly), `medium` (route with clarifying question), `low` (escalate to free-text fallback) This is **Pass 1** of jak.ma's two-pass architecture. Pass 2 (generation + verifier) is a separate model. ## Example inputs | Input | Expe …