Near GPT-4 performance on 60% less of compute for training