Supervised fine-tuning
We fine-tune on examples that pair representative inputs with the outputs we want the model to produce.
Post-training
A model can earn a higher reward and still produce worse outputs. We design training objectives and evaluations together to test whether the model learns the intended behavior and retains the capabilities the application depends on.
Post-training methods
Reinforcement learning
We generate rollouts and score the model’s outputs with a reward function. Those scores guide the next update.
How we evaluate checkpoints
Target behavior
We define what success looks like, then compare checkpoints with the base model and any current production model using the same held-out inputs and settings.
Regressions
We check for regressions on inputs beyond those used to evaluate the target behavior. For video, we check whether scene content, motion or sharpness has changed.
Failure analysis
We inspect outputs, including high-scoring ones, for failures the metrics miss. We document what the evaluation leaves untested.
