Interactive Training 2: A Control Plane for Steering Live Model Training

Most experiment trackers let you watch training happen in real time, but steering an active run—adjusting learning rates, pausing, switching datasets—still demands writing trainer-specific code. Interactive Training 2 solves this by introducing a shared protocol that lets both humans and automated controllers send requests to a training loop using the same interface.

The Core Idea

The system works around three key pieces: a declaration layer, a request interface, and a validation checkpoint. Training applications explicitly declare which settings (hyperparameters, data sources) and actions (pause, checkpoint, resume) they expose. Humans or automated agents then submit control requests through a unified API—no custom code needed. The training loop intercepts these requests at designated safe control points, validates them (checking type, bounds, and state), and applies them if valid.

A customized Aim workspace (Aim is an experiment tracking tool) binds this together visually: users see live metrics and available controls alongside a chronological audit log showing every request, its status, and the outcome. This record proves what changed, when, and why—critical for reproducibility and debugging.

Why It Matters

The practical friction this removes is significant. Today, if you're training a large language model and want to halve the learning rate because validation loss stopped improving, you either write a custom script for that specific training setup or you stop the run entirely and restart from a checkpoint. Interactive Training 2 flattens that workflow. A researcher clicks a slider, an RL experimenter's automated controller sends a JSON request—both work the same way.

The audit trail is equally important. Machine learning practitioners often struggle to remember why they made a change mid-run, especially in multi-day experiments. A chronological log of all interventions and their effects provides explicit accountability and helps teams reproduce results or investigate why a run diverged.

Validation in Practice

The authors demonstrate the system across five workflows spanning NLP and reinforcement learning, showing that the protocol generalizes beyond toy problems. The "safe control points" design prevents race conditions: the training loop only checks for pending requests at defined moments (typically between training steps), so validation logic doesn't need to lock shared state or interrupt computation.

The Larger Picture

Interactive Training 2 treats training as something you can interact with while it runs, not something you launch and observe passively. By enforcing a shared protocol, it eliminates the fragmentation where each training framework or team invents their own steering mechanism. The open-source release of code and real execution traces gives others a tested foundation to build on—whether they want to add human-in-the-loop controls, integrate automated meta-learners, or simply improve debugging practices.

For teams managing many simultaneous experiments or deploying trained agents in adaptive settings, this is infrastructure that reduces both operational friction and the risk of losing critical context about what actually happened during training.