Stripe: fine-tuned 4B open-weight model beats prompting by 24% AP in fraud review
emilygsands · x · 2026-10-15
- Stripe engineers fine-tuned an open-weight 4B model on fraud review labels, reading a fraud score from yes/no token probabilities: average precision rose 24% over prompting, beating human reviewers at the same recall.
- Their thesis: fine-tuning pays off when you need precise evaluations plus rich, heterogeneous evidence.
- Linear probes suggest a key driver is that fine-tuning made legitimate businesses' signals easier to decode.
Related event: Stripe's fine-tuned 4B open model beats prompting by 24% in fraud review(2 posts)→
More from Companies & People
- 85% of US social impact professionals use AI, but only a third say their organizations use it effectively — moniquejmorrow · 2026-10-15
- Someone told Jensen Huang "Europe is done, no AI there" — he clapped back — mattturck · 2026-10-15
- Three questions to ask your AI supplier: capacity ownership, financing, change notice — YvesMulkers · 2026-10-15
- Sergey Brin: Google is already on the recursive self-improvement treadmill, models building models — gaganghotra_ · 2026-10-15
- Pangram CEO on AI detection: model fingerprints, synthetic data and beating humanizers — arena · 2026-10-15
- Vercel's sonofalli departs after ~2 years building its AI ecosystem — sonofalli · 2026-10-15