Stripe Blog Details Two-Stage SFT Plus On-Policy Distillation for Better Fraud Decisions

emilygsands · x · 2026-11-20

The companion Stripe Dot Dev blog post (by Hangzhi Guo) behind the "39% higher recall" tweet: adding explanations directly to SFT hurts decisions because token-uniform loss lets long explanations dominate the gradient. The fix is two-stage SFT—train on labels+explanations first, labels alone second—then on-policy distillation with reverse KL to recover explanation quality. Fraud review achieves 39% higher recall at human-review precision, with explanations beating the teacher's.

Related event: Stripe Boosts Fraud Review Recall 39% by Learning from Explanations(2 posts)→

Original post →

More from coding & agent

coding & agent channel →