Pythia 160M, Mamba 130M, and RWKV 169M models trained on OpenWebText for 4000 steps (context window: 1024; effective batch size: 512). 6 seeds each.