Academic Month Part B

This repository contains a self-contained solution for the Part B prompt-to-SVG logo task. The provided student kit is intentionally not assumed to exist; the reward, LoRA training loop, and self-evaluation script are implemented here.

Environment

The run used a separate conda environment, not care:

conda create -n academic-month-gemma python=3.10 pip
conda activate academic-month-gemma
pip install -r requirements.txt

CUDA is required by the scripts. They intentionally do not fall back to CPU.

Reproduce

Download the base model from ModelScope:

python scripts/download_model.py --model-id google/gemma-3-270m --local-dir ../models/gemma-3-270m

Train the LoRA adapter:

CUDA_VISIBLE_DEVICES=1 python -u -m student_kit.train_peft --config train_config.yaml

Run the self evaluation:

CUDA_VISIBLE_DEVICES=1 python -u -m student_kit.eval_self --config train_config.yaml --output results.json

The submitted adapter is in adapter/, the proxy reward is in reward.py and student_kit/reward.py, and the analysis is in report.md.