mnist-mlp

A small MLP trained from scratch (random init, no pretrained weights) on ylecun/mnist for 5 epochs (256 batch, Adam lr 0.002).

Architecture	784 -> 256 -> 128 -> 10, ReLU
Params	235,146
Steps	1170
Test accuracy (10k)	0.9772
Train time	5.5s on cpu-basic (2 vCPU)

model.pt is a plain state_dict for the class in train.py; reproduce with python train.py.