Towards model-free RL algorithms that scale well with unstructured data