Stripe researchers improve risk models with two-stage SFT

stripe.dev11 min read

Stripe researchers found that training models on risk labels and explanations simultaneously degrades decision performance, but a two-stage supervised fine-tuning approach solves this issue. By training first on labels and explanations together, then on labels alone, the model improves risk assessment recall. Furthermore, applying on-policy distillation as a third stage restores explanation quality without sacrificing the decision accuracy gains. Open article →