Knowledge augmentation is essential to make machine studying fashions extra strong and protected. Nevertheless, augmenting information may be difficult because it requires producing numerous information factors to carefully consider mannequin habits on edge circumstances and mitigate potential harms. Creating high-quality augmentations that cowl these “unknown unknowns” is a time- and creativity-intensive activity. On this work, we introduce Amplio, an interactive instrument to assist practitioners navigate “unknown unknowns” in unstructured textual content datasets and enhance information variety by systematically figuring out empty information areas to discover. Amplio contains three human-in-the-loop information augmentation methods: Increase with Ideas, Increase by Interpolation, and Increase with Massive Language Mannequin. In a person examine with 18 skilled purple teamers, we exhibit the utility of our augmentation strategies in serving to generate high-quality, numerous, and related mannequin security prompts. We discover that Amplio enabled purple teamers to enhance information rapidly and creatively, highlighting the transformative potential of interactive augmentation workflows.
