The company describes it as a group of vetted external companies that will build custom reinforcement-learning environments for Baseten customers undertaking post-training work.
Baseten says it will scope and calibrate an engagement, then refer the customer to a partner. The partner works directly with the customer and its real-world data to create the needed environments. Depending on the engagement, that work can include tasks, task review, rubrics, and resources for producing data. Baseten retains responsibility for model training and production inference.
The company says customers can keep using the resulting environment system rather than treating it as a one-off dataset. It also says it plans to open-source a set of environments, audit methods, and training methods in the future.
The announcement does not name the network partners, give prices, specify customer eligibility, or provide a release date for the open materials. Source: https://www.baseten.co/blog/introducing-the-baseten-environment-network