What We Provide
-
The model(s) you will work with: gemma-4-E2B and gemma-4-E2B-it (instruction-tuned). These are small, open-weight models available on Hugging Face. While you can choose to focus on just one, comparing instruction-tuned vs. non-instruction-tuned behavior is encouraged.
-
Boilerplate code for an LLM-as-judge setup — a scaffold for using a second model to score open-ended outputs.
- Basic Bash Commands.
-
Access to Rivanna — 1 GPU (24GB) per group for the duration of the hackathon.
--account=ds6051-summer
--partition=interactive
--reservation=ds6051-summer-hackathon
--gres=gpu:1
-
Extra GPU access: the free tiers of Google Colab and Kaggle provide T4 GPUs with 16 GB VRAM. This should be enough to run inference on the model, though likely not enough to also run a judge model simultaneously — plan your judge runs accordingly.