Robustness of Learning from Task Instructions