A collection of models, ranging from 255M to 2.8B parameters in size, that are pre-trained on a 272B token data mix containing ObscuraX.