Lab 21 Vietnamese Customer-Support Triage LoRA
LoRA adapter trained on 225 synthetic Vietnamese customer-support tickets to produce a four-field JSON object: intent, urgency, product, and sentiment.
Configuration
- Base:
Qwen/Qwen3.5-9B
- Placement: text-decoder linear modules
- Rank / alpha: 16 / 32
- Learning rate: 1e-4
- Optimizer steps: 30 (2 epochs)
- Precision: bf16
- Hardware: NVIDIA A100-SXM4-40GB
Measured results
- Target field accuracy: 0.990
- Format compliance: 1.000
- Prompted-base target baseline: 0.815
- Regression score: 0.1333 versus 0.7422 for the base
- Regression verdict: FAILED because general capability dropped by 0.609
Limitations
This adapter over-applies the triage JSON schema to general questions. It must not replace a general-purpose assistant without routing or replay-data mitigation. The measured target gain does not offset the general-capability regression for broad deployment.
