A practical checklist before you start post-training

Post-training only pays off when the failure mode is well understood. Before touching weights, we make sure we can answer three questions: what exactly is failing, how often, and what a fixed response actually looks like.

From there, the method mostly picks itself. Narrow, well-specified corrections tend to respond well to SFT on curated examples. Preference-shaped problems — tone, refusals, formatting — respond better to RLHF. Verifiable, checkable tasks are where RLVR earns its cost.

The checklist is short, but skipping any step is the single biggest predictor of a training run that doesn't move the needle.