y-gemma4-svg-lora

QLoRA adapter on unsloth/gemma-4-E4B-it that teaches the model an SVG-native whiteboard tutoring DSL: [text:], [draw_part: name="..." viewBox="..."], raw SVG path-data per line, [/draw_part].

Trained on ~400 rows distilled from duxiaodan/ControlSketch-Part).

This is the Unsloth-track deliverable for the Y project — an AI whiteboard tutor for the Gemma 4 Good Hackathon. See the project repo for the full pipeline.

Usage

from unsloth import FastVisionModel
model, processor = FastVisionModel.from_pretrained(
    "QuantumTransformer/y-gemma4-svg-lora",
    load_in_4bit=True,
)

Training config

• Base: unsloth/gemma-4-E4B-it
• QLoRA: r=16, α=32, dropout=0
• target_modules="all-linear", language layers only (vision/audio frozen)
• 2 epochs, lr=2e-4, batch=2, grad_accum=4
• AdamW-8bit, linear schedule, 10 warmup steps
• Trained on a Kaggle free T4

License

Inherits Gemma 4's license terms.