Notebook fix

[surfiniaburger]

Instead of Print Statements I switched to logger and also saved the logs as the server is booted. So while the server runs we can easily monitor its health. And previously, I was using a reward system that only focused on outcome based scoring, inspired by RLVR I moved to process based scoring. based on a few training I made, I observed that some of the output from the model where malformed, so I moved to "Format-First" Hierarchical Curriculum, where if the model fails the format, we don't consider its reasoning trace. The details to all relevant changes I made are in here.

[pankit-eng]

@surfiniaburger - do you mind adding description on the PR as to what the issue was and what was the fix?

[surfiniaburger]

@pankit-eng I have included the description. I have been undertaking a major refactor, which provides for MCP and A2A. The goal is to build out the pipeline for the DIPG safety Gym as a means to learn more about LLM Alignment and document my findings. It's no small task. Aside from the PR, I'd love to know the areas I should do more research on. Your feedback would be highly appreciated.

[burtenshaw]

@surfiniaburger This looks ok because the changes are purely within the DIPG env. Could you do these to merge:

• merge changes from main.
• push the env to the hf hub.
• updated docs/environments.md with the dipg env and link to the env on the hub

In a future PR, you might also want to add a guide to the docs which explains how to use the dipg env.

[surfiniaburger]

I'll look into it now

[surfiniaburger]

@burtenshaw Thanks for the heads up, I'll work towards adding a guide.