Lab 21 — LoRA adapter for Vietnamese CSKH ticket triage

• Placement: all linear (text-linear), r=16, lora_alpha=32, LR=1e-4
• 2 epochs, 30 steps, effective batch 16, T4 (fp16)
• target accuracy: 0.970 (vs. 0.765 for the best-prompted un-tuned base model)
• Known limitation: the accompanying report found a regression metric drop of 0.147 (catastrophic forgetting on general-purpose questions) that fails the lab's regression gate — see the source repo's submission/REPORT.md for the full writeup, including a systematic bias where the model predicts urgency=trung_binh instead of the correct thap on several held-out examples.