Qwen3.5-4B Real Estate Call Analysis V5 (Egyptian Market)
LoRA fine-tune of Qwen/Qwen3.5-4B for structured-JSON extraction from
Egyptian real estate call transcripts (English + Egyptian Arabic).
Training
| Hyperparameter | Value |
|---|
| Base model | Qwen/Qwen3.5-4B |
| Method | LoRA via Unsloth |
| LoRA rank (r) | 16 |
| LoRA alpha | 16 |
| LoRA dropout | 0.05 |
| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| Max seq length | 2048 |
| Epochs | 6 |
| Effective batch size | 16 |
| Learning rate | 2e-4, cosine schedule |
| Warmup steps | 50 |
| Optimizer | adamw_8bit |
| Gradient clip | 1.0 |
| Weight decay | 0.01 |
| Seed | 3407 |
Dataset
| Item | Value |
|---|
| File | training_data_v10_final.jsonl |
| SHA-256 (first 12) | d2026dc822bf |
| Records | 1026 (train 922, val 104) |
| Train/val split | 90/10 stratified by language, seed=3407 |
Final metrics
| Metric | Value |
|---|
| Train loss | 0.7738 |
| Best clean-eval | 0.6224 |
| Best noisy-eval | 1.4256 |
| Training time | 316.3 min |
Output schema
Structured JSON: client_name, customer_sentiment, urgency, timeline,
confidence_score, transcript_quality_score, client_profile,
special_requests, action_items, call_summary, total_units_requested,
and requested_units[] with intent, property_type, location, currency,
budget ranges, payment_method, area_sqm, bedrooms, finishing, key_objection.
Inference
Quantized to Q4_K_M (~2.7 GB). Run via llama.cpp / llama-cpp-python.
Designed to run on GPUs as small as 6 GB VRAM (RTX 3050 class).
