Would also love this to include InsturctGPT architecture with its RL reward model!