Practical Tradeoffs: Data, Compute, Stability and Real Examples

~12 min read

Beyond the decision tree, SFT and RFT differ practically in data requirements, compute cost, and training stability — tradeoffs worth understanding even after the tree points you toward one.

Practical Tradeoffs: Data, Compute, Stability and Real Examples is a Pro topic

Sign in, then upgrade to Pro or Power to unlock this topic and the full AI Engineering curriculum.

Key points

  • SFT needs a dataset of pre-written correct completions (expensive to source: human annotation or synthetic pipelines like Distilabel)
  • RFT needs a RELIABLE reward function instead — easier for verifiable tasks (check a math answer), hard/impossible for subjective ones
  • SFT costs one forward pass per example; RFT (via GRPO) must generate multiple candidate completions per prompt before it can even compute a reward, costing meaningfully more compute
  • SFT training is stable and predictable (fixed targets); RFT is more prone to instability, including reward hacking (gaming the reward function without solving the real task)
  • The book's own GRPO reward functions (format-exact, format-approximate, answer-check, number-check) are deliberately deterministic and hard to game — RFT's extra cost is worth it specifically when you can design a reward this reliable