Is there a comparison of training speed with the implementation in TWOSOME?

[elggurts22]

Is there a comparison of training speed with the implementation in TWOSOME?

[KhoomeiK]

Haven't looked too closely at TWOSOME, but Lamorel is definitely faster and more efficient if you want to train large models in distributed compute setups. LlamaGym is meant to be a minimalistic, quick setup, and easy to understand introduction to running online RL experiments with LLM agents.