Training a local model to play browser games on the fly

An interesting approach where a frontier model (Claude) is used to train the local model (Laya)to play browser games and learn as it fails. Both Jev and Laya were tried, and Laya was preferred due to need for low latency.

How do you train your model? Does it learn through self-play?

IronBee Gamer author here.

An LLM writes each game's rules as code. The rules play the game and label every state with the move they would make. The model (322M, mmBERT-base) is fine-tuned on those labels. Then the model plays its own games, the rules label the states it ended up in and it's fine-tuned again.

Thank you . How do you prevent the model from overfitting to the rule-based labels, and can it eventually discover strategies that outperform the original rules?

sick