google/dopamine ? reverse-engineered prompt

Build me a small research framework for deep reinforcement learning that’s easy to understand and experiment with.

I want it to support Atari and Mujoco environments, with a clean way to train and evaluate a few standard agents like DQN, C51, Rainbow, IQN, SAC, and PPO. Keep the codebase compact and readable, and make sure it’s set up so new ideas can be added without a lot of friction. If it makes sense, use JAX for the main implementations, and keep any older TensorFlow support separate if needed.

Please include a simple way to run benchmark experiments, some example notebooks or quick start scripts, basic tests, and docs that show how to install dependencies, run training, and reproduce results. If there are obvious best practices for Docker or virtual environments, set those up too. I’d also like baseline results or example outputs so I can tell everything is working. Look up current docs online if you need to.