TL;DR

I bring you:

• Sandboxing environment to train shell agents: repo
• A 14k shell task dataset for training on HF
• Synthetic data generation pipeline to generate customizable task datasets
• Batteries-included script to run RL with the above and your model of choice

Intro

Earlier this year I started messing more with “agents”, so much more that it ended up changing what I do at work (cool).