Training on explanations helps Stripe make more precise risk assessments. We explain our two-stage SFT approach and how on-policy distillation restores the model’s ability to explain its decisions.
SFT helps Stripes make more precise risk assessments. We explain the training approach and investigate what the model learns about legitimate businesses.
Authors:
Hangzhi Guo, Stathis Vafeias, Florian Hartl, Luvsandondov Lkhamsuren, Adrian Lam