QunAnh/Lab21-LoRA-CSKH

Lab21 - Qwen3.5-4B LoRA for Vietnamese customer-support triage

This is the correct text-decoder LoRA adapter from Lab21. Training used 225 synthetic Vietnamese customer-support tickets, rank 16, alpha 32, learning rate 1e-4, assistant-only labels, max_length 256, two epochs, and 30 optimizer steps on Tesla T4 with fp16 computation.

Evaluation

Full frozen evaluation: 50 target tickets and 15 general-capability prompts. Target is mean four-field accuracy; regression is keyword recall. Greedy generation uses the same lab harness.

run	target	regression	format	latency_ms	n
(a) base + naive prompt	0.0	0.7911	0.0	3648.9	50
(b) base + optimized prompt	0.765	0.7911	1.0	1124.4	50
(c) LoRA fine-tune	0.97	0.4778	1.0	1696.4	50

Regression gate verdict: FAIL. Reasons: general capability regressed by 0.313 (tolerance 0.020). See deck §6.3 — add 1-5% replay data.. This is a lab artifact; the evaluation does not establish readiness for production or general assistant replacement. See submission/REPORT.md and results/ for all evidence, qualitative errors, and trade-offs.

Loading

Load base checkpoint unsloth/Qwen3.5-4B with the corresponding tokenizer, then load this repository with PeftModel.from_pretrained. Use system prompt "Phân loại ticket sau." and greedy decoding, with the model's chat template and thinking disabled as in labkit.generate. This repository contains an adapter rather than a merged base model.

Submission

GitHub source: https://github.com/AnhQun18/Day21-Track3-Finetuning-Lab Hugging Face: https://huggingface.co/QunAnh/Lab21-LoRA-CSKH

No custom corpus, NB6, reasoning-trace ablation, or rank sweep bonus is claimed.