LoRA adapter for Qwen3-VL-4B-Instruct fine-tuned to answer natural-language questions about Wikipedia-screenshot chunks, specifically on images compressed by 5x (each dim scaled by 1/√5).

Performance (GPT-4.1 LLM-judge on 500 test examples)

Setup	LLM-judge
Uncompressed (0x) ceiling, base Qwen3-VL-4B	0.958
This adapter @ 5x	0.730
Base Qwen3-VL-4B @ 5x (no SFT)	0.554

SFT gain over base at 5x: +0.176 (31.8% relative).

Training config

Method: LoRA (r=128, 2ep, lr 3e-5, ViT unfrozen)
Base model: Qwen/Qwen3-VL-4B-Instruct
Checkpoint: step 2400
Framework: LLaMA-Factory (fork)
Trained on 4× H100 80GB (DeepSpeed ZeRO-2, bf16)
Dataset: Wikipedia screenshot-QA pairs compressed with PIL LANCZOS

Data preparation

Training images were downscaled by 1/sqrt(5) per dimension using PIL LANCZOS, e.g. a 1200×800 screenshot becomes 536×357 px (~20% of original pixels).

Usage

python

from transformers import AutoProcessor, Qwen3VLForConditionalGeneration
from peft import PeftModel
import torch

base = "Qwen/Qwen3-VL-4B-Instruct"
adapter = "Chrisyichuan/qwen3vl-4b-wiki-screenshot-5x-lora"

model = Qwen3VLForConditionalGeneration.from_pretrained(base, torch_dtype=torch.bfloat16).cuda()
model = PeftModel.from_pretrained(model, adapter).merge_and_unload()
processor = AutoProcessor.from_pretrained(base)

# PIL image already compressed to 5x
messages = [{"role": "user", "content": [
    {"type": "image", "image": your_compressed_image},
    {"type": "text",  "text": your_question},
]}]
# ... standard Qwen3-VL inference

Notes / limitations

The adapter is specific to the 5x compression level and does not necessarily generalize to higher or lower compression. Use the adapter whose level matches your deployment.
At 5x, SFT recovers 44% of the compression-induced accuracy drop relative to uncompressed Qwen3-VL-4B.
See the full experiment matrix and findings in sft/RESULTS.md of the source repo.