Training data
Training data, in the interface-behavior context, is recorded interaction data used to teach models how software is operated: demonstrations of tasks, successful agent trajectories, and human sessions that show real usage patterns. Session recordings are a natural source, provided consent, privacy processing, and format conversion are handled deliberately.
How it works
The pipeline runs from recordings to examples: select useful sessions, remove personal data beyond what capture-time masking handled, and convert the event stream into the model's observation-action format. DOM event streams already contain structured page changes and element-referenced interactions, while video requires another system to recover that structure. Training is a separate processing purpose that needs its own legal basis and governance. Trajectories produced in a controlled synthetic environment can reduce reliance on recordings of people, but agent-generated data is not automatically exempt. A run can still contain personal data, confidential account content, or material from third-party pages.
Where you encounter it
Agent fine-tuning, imitation learning from demonstrations, and the pipeline design for turning user sessions into training data.