---
license: mit
tags:
- mnist
- learning-rate-sweep
- from-scratch
---
# MNIST CNN from scratch — SGD lr=0.003

Trained for 60s wall clock (batch 128, SGD momentum 0.9, seed 0), reaching step 10426 with final train loss 0.001416308805346489 and test accuracy **0.9879**.

Part of a 10-learning-rate sweep. Live curves: https://huggingface.co/spaces/abidlabs/mnist-lr-sweep-trackio