Everything needed to develop a solution to this task and to score it the way the contest did.
code/
baseline-original/ the baseline exactly as kept during the contest
grading-original/ the grader exactly as used to score submissions
The data is not in this repository — it is published as a standalone Hugging Face dataset:
The dataset has public/ (what contestants received) and private/ (the hidden leaderboard A/B configurations and held-out test) — its dataset card documents every subset.
Download it and place it so the notebook can find it (see the note below), then open the baseline.
Running the baseline
The baseline trains a small network from scratch to reproduce the logo field, then saves it as model.pt.
cd code/baseline-original
PYTHONPATH=../grading-original jupyter lab solution.ipynb
custom_model.py holds the CustomModel class the notebook trains.
The notebook imports core, which lives in ../grading-original/core — during the contest it was mounted read-only next to the notebook. Set PYTHONPATH=../grading-original (as above) or the import will fail. The model must stay under 20,260 parameters or the score is halved.
Scoring a solution
code/grading-original/ holds the grader used during the contest. It reads a solution plus the private data and prints a score. The grader's own README documents how it is invoked and what it expects; it was built to run inside a prepared container image, so running it locally takes some setup.
A note on the baselines
These are deliberately simple — several score close to zero. They are starting points that show the input and output format correctly, not strong solutions. Improving on them is the exercise.