Catalog

When training externally facing large language models (LLMs) on internal or sensitive data, you need to be sure that you don’t unintentionally or inappropriately disclose portions of the training data. In order to protect against sensitive data disclosure, the general recommendation is to not train these complex systems with any non-public customer data. However, this may not always be possible. This chalk talk addresses how to mitigate this risk. Explore data minimization principles such as sanitization, anonymization, synthetic data generation, aggregation, and differential privacy while also discussing the benefits and tradeoffs.