LoRA adaptation of Qwen3-VL-2B-Instruct for front-view autonomous-driving visual question answering on the Bench2Drive (Chat-B2D) benchmark. The vision encoder is frozen; LoRA is applied to the language model + multimodal aligner, then merged back into a single standalone checkpoint.

Input format: front camera current frame + 3 history frames (oldest→newest), with a driving-agent system prompt. Trained on driving data only (no general-domain replay) — this recipe was found to outperform general-mixed and data-scaled variants on both driving and general benchmarks.

Results (held-out front-view QA, n=2,500, deterministic split, zero clip overlap)

Model	Recipe	BLEU-4	ROUGE-L	CIDEr	MMStar
Qwen3-VL-2B (stock)	—	3.00	14.15	0.47	43.67
This model (v7)	B2D-only, 56k, 2 epochs	45.40	74.30	436.4	52.60

General-ability retention (same scorers): AI2D 71.6 · ChartQA 77.2 · RealWorldQA 62.9 · ScienceQA 79.4 · MMStar 52.6 (stock MMStar 43.7 — general reasoning is preserved and, on several benchmarks, improved).

Usage

python

from transformers import AutoModelForImageTextToText, AutoProcessor
m = AutoModelForImageTextToText.from_pretrained("HqH1111/Qwen3-VL-2B-Bench2Drive-FrontView", torch_dtype="auto", device_map="auto")
p = AutoProcessor.from_pretrained("HqH1111/Qwen3-VL-2B-Bench2Drive-FrontView")

Training

Base Qwen3-VL-2B-Instruct · ViT frozen · LoRA r64/α128 on LLM + aligner · bf16 · 3×A100 · ms-swift. Driving metrics via pycocoevalcap (BLEU-4/ROUGE-L/CIDEr). Non-commercial research use.