Anakonkai/qwen3.5-9b-lora-traffic-v2
010
qwen3.5-9b-lora-traffic-v2
QLoRA adapter fine-tuned on Qwen/Qwen3.5-9B for Vietnamese traffic-law QA.
This is the canonical adapter for configs C (no RAG) and D (with RAG).
Training
- Base:
Qwen/Qwen3.5-9B, 4-bit NF4 via unsloth. - LoRA rank 32, alpha 64, targeting all 7 modules (q/k/v/o, gate/up/down_proj).
- Data:
data/splits_filtered/qa_train.jsonl(1762 pairs) + dev 220. - Schedule: 2 epochs, lr 5e-5, batch 2, grad_accum 8 (eff batch 16), cosine warmup 5%.
CONTEXT_KEEP_PROB=0.9— 90% of samples keep the raw legal context in the prompt.
Usage
from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
model_name="Anakonkai/qwen3.5-9b-lora-traffic-v2",
max_seq_length=2048, load_in_4bit=True,
)
FastLanguageModel.for_inference(model)Results (145-sample labelled eval)
- Config C (LoRA, no RAG): ROUGE-L 0.389, BLEU 0.146, BERTScore 0.638, LLM-Judge 0.39.
- Config D (LoRA + RAG): ROUGE-L 0.515, BLEU 0.350, BERTScore 0.692, LLM-Judge 0.69.
