Experiments

The Experiments panel tracks every training run in your project. Each experiment is logged as a row in experiment_log.tsv with its configuration, metrics, and outcome.

Key Capabilities

• Automatic logging -- The /experiment cursor command appends to the log after each run
• Side-by-side comparison -- Compare hyperparameters and metrics across any two experiments
• Status tracking -- Each run is marked as success, failure, or partial
• Metric history -- View how loss, accuracy, or any custom metric trends over time
• Configuration diffs -- See exactly what changed between two runs

How It Reads Data

The Experiments panel reads from:

• experiment_log.tsv -- The primary data source. Each row contains a timestamp, experiment ID, status, description, and JSON-encoded metrics and configuration.

The dashboard parses this file on startup and watches it for appended rows. New experiments appear in the UI within seconds of being logged.

Framework-agnostic

ResearchPad does not hook into your training framework. The cursor commands write to experiment_log.tsv using plain text. This means it works with PyTorch, TensorFlow, JAX, scikit-learn, or any other library.