model-soups

Why we included this project

If you are training several fine-tuned models of the same architecture and want the accuracy gains without paying for multiple inference passes, this repo is a practical starting point. It implements the two main soup approaches from the paper: a uniform soup that averages all model weights, and a greedy variant that only adds a model when it improves held-out accuracy. The workflow is broken into separate commands for downloading, evaluating, running each soup, and plotting, so you can rerun just the parts you care about. A linked Colab notebook walks through the same idea on CIFAR-10, which is a quick way to see whether the technique.