Educational LoRA adapter from a fine-tuning lab (AICB course, Day 21). Trained to convert Vietnamese customer-support tickets into a 4-field JSON triage (intent, urgency, product, sentiment).

Full report, all raw results, and the training/eval pipeline: https://github.com/phamquoctuan2308/Day21-Track3-2a202601983-PhamQuocTuan

⚠️ Known issue — do not use as a general-purpose assistant

This adapter was trained for 2 epochs on 225 narrow-domain examples with no replay data. It scores excellently on the target task but has catastrophic forgetting of general capability:

Metric	Base + optimized prompt	This adapter
Target task accuracy	0.595	0.995
General-knowledge regression score	0.689	0.067
JSON format validity	1.000	1.000

11/15 general-knowledge questions in the regression eval get a JSON triage object back instead of an answer — e.g. asking "Ai là tác giả của Truyện Kiều?" ("Who wrote Truyện Kiều?") returns {"intent": "hoi", "urgency": "trung_binh", ...} instead of "Nguyễn Du". The model has effectively lost the ability to do anything except emit this JSON schema. See the linked report §5–6 for the full regression-gate analysis (verdict: FAILED) and the fix that was not applied here (1–5% general-purpose replay data mixed into training, per LoRA-fine-tuning best practice).

Training config

Base model: Qwen/Qwen3.5-2B
LoRA: r=16, alpha=32, target modules = all linear layers (text-linear placement), 16,819,200 trainable params
Learning rate: 1e-4 (cosine schedule), 2 epochs, 58 optimizer steps
Precision: bf16, effective batch size 8 (per_device_batch=1 × grad_accum=8)
Mask mode: assistant-only (loss computed only on the JSON answer span, supervised_fraction=0.394 — verified by decoding the supervised span, not assumed)
Hardware: RTX 3060 Laptop 6GB, ~36 min wall-clock

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-2B", dtype="bfloat16")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-2B")
model = PeftModel.from_pretrained(base, "TuanPham2308/lab21-qwen3.5-2b-cskh-triage-lora")

Framework: PEFT 0.20.0