Comparison of downsampling methods after 2.5B tokens