Lab21 Qwen3.5-4B LoRA adapter
Public adapter for VinUni Track 3 Lab21, submitted by Nguyễn Vũ Anh (2A202602502).
This repository contains the trained correct PEFT adapter, not merged base weights. Load with unsloth/Qwen3.5-4B. Training used 225 synthetic Vietnamese customer-support tickets (250 total, 25 validation), assistant-only loss masking, text-linear LoRA rank 16 / alpha 32, LR 1e-4, sequence length 256, seed 42, fp16 on Tesla T4, 30 scheduler steps.
Evaluation and limitations
Full evaluation: 50 target tickets and 15 regression questions. Target field accuracy 0.98 (46/50 tickets entirely correct); format score 1.0; regression keyword recall 0.522222; latency 1352.4 ms/item. Frozen optimized base baseline: target 0.765, regression 0.791111.
Experiment verdict: FAIL because regression fell by 0.268889, exceeding the 0.02 tolerance. The adapter can over-apply the ticket JSON schema to general questions. This is an educational experiment, not validated for general assistant deployment. fp16 logs contained intermittent NaN gradient norms; saved adapter weights were verified finite. Actual skipped GradScaler updates were not instrumented.
Full report and source · Report · Results
The base model's applicable terms continue to apply; this upload contains only adapter and tokenizer assets.
- Downloads last month
- 20