mahmoudd777/qwen35-realestate-gguf-v3

Qwen3.5-3B Real Estate Call Analysis V3 (Egyptian Market)

Fine-tuned Qwen3.5-3B-Instruct for extracting structured JSON from Egyptian real estate call transcripts. Supports English and Egyptian Arabic (colloquial).

What it does

Given a call transcript, extracts a structured JSON with:

• —Client name, sentiment, urgency, timeline
• —Confidence score, transcript quality score
• —Client profile, special requests, action items, call summary
• —requested_units: intent, property type, location, currency, budget, payment method, area, bedrooms, finishing, key objection

Training (V3)

• —Base model: Qwen/Qwen3.5-3B-Instruct
• —Method: LoRA fine-tuning (r=16, alpha=16) via Unsloth
• —Dataset: 850 records (57% English, 43% Egyptian Arabic)
• —Train/Val split: 90/10 with seed=42
• —Sequence length: 2048 tokens
• —Epochs: 8

V3 Improvements over V2

• —+70 more records (780 -> 850)
• —Better Arabic coverage: 36% -> 43%
• —Targeted fixes for confirmed model weaknesses:
• —October location: client says the name 3+ times to prevent confusion
• —Nasr_City: explicit "مدينة نصر" to stop Heliopolis confusion
• —Client/agent name: agent name appears once, client name 3+ times
• —Added Zamalek, Alexandria, Giza records (previously underrepresented)
• —Added validation split to detect overfitting
• —Increased epochs: 6 -> 8 for better convergence