You can deploy this model and start earning money today!

tuan2294/
lab21-2A202601618-qwen35-triage-vi

Name: lab21-2A202601618-qwen35-triage-vi
Availability: InStock

Views

No views yet

Lab 21 — LoRA adapter · Qwen3.5-9B · Vietnamese CSKH ticket triage

LoRA adapter (r=16, alpha=32, target_modules=text-linear, LR 1e-4, 2 epochs / 30 steps) fine-tuned on Qwen/Qwen3.5-9B for the task: Vietnamese e-commerce support ticket -> structured JSON triage (intent, urgency, product, sentiment).

Full methodology, code, and all results/*.json are in the companion GitHub repo: https://github.com/nt15032/Day21-Track3-Finetuning-Lab-2A202601618-NguyenDucAnhTuan

Full report: submission/REPORT.md in this repo (mirrored from the GitHub repo above).

Headline result

	target	regression	format	latency (ms)
(a) base + naive prompt	0.000	0.742	0.000	2878
(b) base + optimized prompt	0.815	0.742	1.000	767
this adapter	0.990	0.133	1.000	1239

Verdict: FAILED the four-group regression gate — the adapter beats the strong prompted baseline on the target task (+0.175) but collapses general capability (regression 0.742 -> 0.133), a textbook catastrophic-forgetting case caused by training on a single-shape corpus (bare ticket -> JSON) with no replay data. See submission/REPORT.md section 5 for the full diagnosis. Not recommended to deploy as-is — needs 1-5% replay data mixed into training before it is safe as a general-purpose endpoint.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-9B", dtype="auto", device_map="auto")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-9B")
model = PeftModel.from_pretrained(base, "tuan2294/lab21-2A202601618-qwen35-triage-vi")

messages = [
    {"role": "system", "content": "Phân loại ticket sau."},
    {"role": "user", "content": "Shop ơi, đơn DH123456 giao chậm quá, đã 3 ngày rồi."},
]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=160, do_sample=False)
print(tok.decode(out[0, inputs.shape[-1]:], skip_special_tokens=True))

Files

adapter_model.safetensors, adapter_config.json — the LoRA adapter
results/ — all NB1-NB6 artifacts (mask proof, frozen baselines, runs.csv, verdict, autopsy of 3 misconfiguration contrasts, qualitative examples, merge check)
submission/REPORT.md — full write-up