IT company SQUAD will host AI meetup “How to make the right to be forgotten a reality with LLMs?”

A meetup for those who research, fine-tune, or deploy LLMs: how to distinguish genuine forgetting from merely hiding information, how to verify the results, and when unlearning is worth applying despite its limitations.

Large language models can memorize and reproduce personal data, confidential information, and fragments of copyrighted content. Simply removing documents from the training dataset does not solve this problem, while retraining a model from scratch can be costly and impractical.

On November 19 in Lviv, SQUAD’s AI specialists will explore how LLM unlearning aims to address this challenge — and why implementing the “right to be forgotten” in language models is more complicated than it may seem.

What’s on the agenda:

• Two approaches to approximate unlearning: gradient ascent and task-vector arithmetic;
• What happens after “forgetting”: can information still be retrieved by rephrasing a prompt? Can it resurface after fine-tuning or quantization? And does the model lose useful capabilities along with the forgotten information?
• Differential privacy, using VaultGemma as an example, and the trade-offs between privacy and model performance;
• Q&A session with SQUAD’s AI specialists — get your questions answered live.

Join us for insightful discussions and engaging networking!

When and where?

• November 19, 7:00 PM
• Lviv, SQUAD’s safe space
• Online stream for those who can’t join in person
• Free admission

Registration is required.