Sentient launches open Fingerprinting Leaderboard for scoring model provenance methods

Sentient has launched the Fingerprinting Leaderboard, a free, open-source benchmark that scores methods for hiding ownership marks in AI models, according to Elvis Saravia and Rohan Paul's reports on X. Each method is measured on how much model quality it preserves, how well it survives removal attempts such as fine-tuning and distillation, and how reliably the owner can detect it.

The benchmark embeds each submitted mark in Sentient's test models, tries to strip it out, and ranks methods on a public site. Sentient released the mlprints code that runs the tests. Sentient says the aim is to keep proof of provenance in the model weights rather than in documentation.

Paul reports that Sentient's Perinucleus method embedded 24,576 fingerprints in Llama-3.1-8B without degrading utility and survived standard fine-tuning. The reports do not give results for any other method.