Divit56/VLM_grader

VLM Art Grader - LoRA Adapters

Fine-tuned LoRA adapters for automatic grading of children's artwork on 3 rubric dimensions: Clarity, Detail, and Creativity. Built to reduce API evaluation costs to under ₹0.10 per image.

Model Details

• Base Model: Qwen/Qwen2-VL-2B-Instruct
• Method: LoRA (Low-Rank Adaptation)
• Training: 3 epochs on 4,000 labeled children's drawings enriched with Chain-of-Thought reasoning.
• Hardware: Single NVIDIA Tesla T4 GPU

Performance

• Mean Absolute Error (MAE): 0.247
• Classification Accuracy: 87.3%
• JSON Parse Success Rate: 100%

Usage

from peft import PeftModel
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor

# Load the base model
base_model = Qwen2VLForConditionalGeneration.from_pretrained(
    "Qwen/Qwen2-VL-2B-Instruct",
    load_in_4bit=True,
    device_map="auto"
)

# Load these LoRA adapters
model = PeftModel.from_pretrained(base_model, "Divit56/VLM_grader")