These last few months I've been working with the RLVR/PPO/GRPO side of post-training. It is a very fun area with lots of material to start from (papers, blogposts, repos), but where a common theme is that most of the challenges the authors faced are "solved". The messy part is not really shown: building a harness where a model is able to interact correctly and get a learning signal, tuning the hyperparams, and iterating a lot over the unexpected errors and pitfalls of RL. And when it is shown, it comes already solved.

I wanted to learn by facing those challenges and solving them myself, so I thought "what's an area where LLMs could learn from a harness, but no one has done it before?" and of course, MTG came to mind.

This is an ongoing project (currently training a 4B SFT model). Expect more info soon.