Questions about Verl training process

[Skyorca]

Hello, may I ask which version of Verl was used for training? I noticed that the inference process uses a continuation-based approach, as shown in the /src/run_deep_agent.py. However, the Verl 0.7 training framework, specifically the agent_loop mode, only supports multi-turn dialogue format for inference.

[sunnynexus]

Thanks for your interest! We've made some minor adjustments to the verl training code and its support for each environment, and did not use the vanilla verl directly.