afifaimran/qwen25vl3b-scienceqa-lora-textonly

Qwen2.5-VL-3B · ScienceQA (text-only LoRA)

LoRA adapter trained on text-only prompts for ScienceQA and evaluated with images at inference.

Quick use

from transformers import AutoProcessor, AutoModelForCausalLM
from peft import PeftModel
import torch
BASE = "Qwen/Qwen2.5-VL-3B-Instruct"
ADAPTER_REPO = "afifaimran/qwen25vl3b-scienceqa-lora-textonly"
processor = AutoProcessor.from_pretrained(BASE, trust_remote_code=True)
base = AutoModelForCausalLM.from_pretrained(
    BASE, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True
)
model = PeftModel.from_pretrained(base, ADAPTER_REPO)  # loads 'adapter' folder
model.eval()

Results (images if available)

See eval/scienceqa_report.md and CSVs:

• eval/scienceqa_qwen25vl_results.csv
• eval/scienceqa_qwen25vl_sc_sweep.csv

Best setting from our sweep: Self-consistency K=5, T=0.7. Last updated: 2025-08-23 08:24 UTC