This talk is part of a two-presentation session running from 2:50 PM - 3:40 PM. The session will feature two presentations back-to-back, with Q&A after each presentation.

When training AI models to detect harmful content on a platform, the quality and volume of training data makes a huge difference in the model's performance. Finding large volumes of rarer kinds of abusive images or texts can be challenging, which can hurt a model’s ability to detect these categories. By fine-tuning generative models to create new examples of text or images with these classes of abuse, we can improve data quality, while also reducing labelling costs and protecting user privacy. In this presentation, we will cover:

• Why you might choose to use generative models over other data augmentation techniques.
• How to fine-tune a generative model, including:
  • How to pick which foundational model to use.
  • Selecting data to fine-tune. Choosing relevant prompts.
  • Overcoming safeguards in generative models.
• The main benefits and drawbacks of this approach.
• The ethical and legal concerns this brings.