Qwen3.5-4B LoRA — Vietnamese Customer Support Triage
LoRA adapter cho bài toán phân loại ticket chăm sóc khách hàng tiếng Việt
thành JSON gồm bốn trường: intent, urgency, product, sentiment.
Base model
unsloth/Qwen3.5-4B
Training configuration
- LoRA rank: 16
- LoRA alpha: 32
- Learning rate: 1e-4
- Optimizer steps: 30
- Mask mode: assistant-only
- Target placement: text-linear
Evaluation
- Fine-tune target: 0.970
- Optimized-base target: 0.765
- Fine-tune format: 1.000
- Fine-tune regression: 0.5667
- Optimized-base regression: 0.7578
The adapter improved target accuracy but failed the regression gate because
general capability decreased by 0.1911. See submission/REPORT.md for the
complete analysis.
This is a PEFT adapter, not a standalone base model.
