learn about how we post-train
castform is a post-training platform that enables you to fine tune models with reinforcement learning. while we manage the training infrastructure and synthetic pipelines to make it easy for you, it is helpful to know the fundamentals. below are some guides on reinforcement learning, fine-tuning, and the ideas behind modern, custom ai systems.
what is an rl environment?
a plain-language guide to what rl environments actually are, what's inside one, and why so much frontier model work comes down to who can build better ones
grpo explained: group relative policy optimization for llm finetuning
a quick intro to grpo - a simple rl algorithm for fine-tuning llms on verifiable tasks
reward hacking: when your ai aces the wrong test
reward hacking is what happens when an ai finds an unintended shortcut to score well on its reward function - without actually doing what you wanted. here is how it works, why it matters, and why fixing it is harder than it sounds.
the state of continual learning
most AI systems are trained once and frozen. continual learning is the project of making them get better over time - and the answer looks different depending on where in the stack you look.