Stripe's fine-tuned 4B open model beats prompting by 24% in fraud review

Stripe's engineering team fine-tuned an open-source 4B model for merchant fraud review, reading fraud scores from yes/no token probabilities. The approach improved average precision by 24% over prompting and exceeded human reviewers at equal recall.

2026-10-15 ~ 2026-10-15 · 2 related posts

• Stripe: fine-tuned 4B open-weight model beats prompting by 24% AP in fraud review — emilygsands · 2026-10-15
• Stripe Finds Fine-Tuned 4B Open-Weight Model Beats Prompting by 24% AP on Fraud Review — emilygsands · 2026-10-15