The secret is to basically use RL to create a model that will generate synthetic data. Then you use the synthetic dataset to fine-tune a pretrained model. The secret is basically synthetic data imo: https://medium.com/thoughts-on-machine-learning/the-laymans-...