Sarus: LLM’s and Privacy
Sarus is one of five winners of the first ELSA Industry Call and has received 60,000 euros in funding, working together with top AI experts within our Network of Excellence for six months. Sarus solves the problem of accessing or sharing personal data for analytics or ML. Every interaction with sensitive data is protected with the highest privacy standard: differential privacy.
In this video, Sarus presents the joint project with ELSA:
Sarus’ short Summary of the project:
“During the program, Sarus built Sarus AI: a module of Sarus App that enables potentially untrusted data-practitioners to fine-tune a LLM such as Mistral, Llama3 or Phi3 on privacy sensitive datasets with differential privacy guarantees on the resulting model.
Because Sarus AI is integrated with Sarus App, the data-practitioners can interactively build a data pipeline to format the sensitive input as text-completion (or chat-completion) training examples suitable for LLM fine-tuning, without ever exposing the original data.”