Chrisyichuan/qwen3vl-4b-wiki-screenshot-2x-lora

Qwen3-VL-4B Wikipedia Screenshot QA LoRA — 2x compression

LoRA adapter for Qwen3-VL-4B-Instruct fine-tuned to answer natural-language questions about Wikipedia-screenshot chunks, specifically on images compressed by 2x (each dim scaled by 1/√2).

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

SFT gain over base at 2x: +0.044 (4.9% relative).

Training config

• —Method: LoRA (r=128, 2ep, lr 2e-5)
• —Base model: Qwen/Qwen3-VL-4B-Instruct
• —Checkpoint: step 2000
• —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(2) per dimension using PIL LANCZOS, e.g. a 1200×800 screenshot becomes 848×565 px (~50% of original pixels).

Usage

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

base = "Qwen/Qwen3-VL-4B-Instruct"
adapter = "Chrisyichuan/qwen3vl-4b-wiki-screenshot-2x-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 2x
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 2x compression level and does not necessarily generalize to higher or lower compression. Use the adapter whose level matches your deployment.
• —At 2x, SFT recovers 81% 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.