mnist-cnn-scratch-1min

A small CNN trained from scratch on MNIST in ~60 seconds of wall-clock training on cpu-basic (2 vCPU), reaching 98.31% test accuracy in 911 steps (batch 128, Adam lr=1e-3).

Architecture: Conv(1→32, 3×3) → ReLU → MaxPool → Conv(32→64, 3×3) → ReLU → MaxPool → Flatten → Linear(3136→128) → ReLU → Linear(128→10).

Training loss curve: https://huggingface.co/spaces/abidlabs/mnist-cnn-scratch-1min-trackio

Run logs: https://huggingface.co/jobs/abidlabs/6ab594d26b030d633f68f6c0