Interactive training with reasoning-gym server

[andreaskoepf]

Vision:
Launch a training run and use cli-commands (or a web-frontend) to monitor and manipulate the reasoning-gym dataset configuration - to directly control the next batch composition, e.g. add or remove datasets from a Composite or change the difficulty: Tune configuration parameters and immediately see the response in the current training or eval run. Work towards a vision for LLM training where humans oversee the evolving training of an LLM during RL - steering the development in the desired direction.

Implementation sketch:

• expose REST API (accessible via API-key)
• offer endpoints to read & manipulate the active configuration & score board
• allow running reasoning-gym parameter server stand-alone (for multi-proc & distributed training with central reasoning-gym server)
• add client class to fetch next task & return results to reasoning-gym server
• create console-client app to read & edit configuration params & monitor accuracy values and current capabilities