MNIST Challenge.
Build. Submit. Iterate.
Teams train an MLP of at most 204,800 parameters, export it as model.onnx, and submit it from their repository. Every submission passes the same contract as evaluate.py and is scored on hand-drawn digits: the public half shows up here right away; the hidden half decides the final ranking. Depth, originality and documentation count too.
Restriction
MLP only
≤ 204,800 parameters, no convolutions
Format
ONNX
uint8 image in, float32 logits out
Attempts
10 / day
per team, reset at 00:00 UTC