diffusion-data-constraint. Official PyTorch implementation and models for paper "Diffusion Beats Autoregressive in Data-Constrained Settings". We find diffusion models are significantly more data-efficient than standard left to right autoregressive models, due to their ability to learn from different token orderings.