Hugging Face has released Interactive Training 2, an upgraded open-source control plane designed to enable live steering of machine learning training runs. Unlike traditional experiment trackers that only monitor progress, this system allows both humans and automated agents to modify optimization parameters and gradients during training through a standardized interface.
The core innovation is a shared protocol where training applications declare which settings and actions they expose. Controllers—whether human or automated—submit requests through the same interface, and the training loop validates and applies them at safe control points. A customized Aim workspace combines live metrics and controls with a chronological record of all requests and outcomes, ensuring full auditability.
The system has been demonstrated across five NLP and reinforcement-learning workflows. The released code and traces provide a reusable foundation for auditable human- and agent-guided training. A live demo is available at interactivetraining.ai/live.