mahmoudd777/qwen35-realestate-2048_V2

Qwen2.5-3B — Real Estate Call Analysis V2 (Egyptian Market)

Fine-tuned Qwen2.5-3B-Instruct for extracting structured JSON from Egyptian real estate call transcripts. Supports English and Egyptian Arabic (colloquial).

What it does

Given a call transcript, the model extracts a structured JSON object with:

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

Training (V2)

• —Base model: Qwen/Qwen2.5-3B-Instruct
• —Method: LoRA fine-tuning (r=16, alpha=16) via Unsloth
• —Dataset: 780 real estate call transcripts (61% English, 39% Egyptian Arabic)
• —Sequence length: 2048 tokens
• —Epochs: 6

V2 Improvements over V1

• —+130 more training records (650 -> 780)
• —Better Arabic coverage: 30% -> 39%
• —New patterns: location disambiguation, frustrated clients, first-time buyers, corporate buyers, zero-result calls, USD currency, calling-on-behalf scenarios
• —Doubled context window: 1024 -> 2048 tokens
• —More diverse locations: Alexandria, Giza, Zamalek, North Coast sub-areas

Language support

• —English call transcripts
• —Egyptian Arabic call transcripts (colloquial)
• —Mixed Arabic/English conversations