Skip to main content

Replicability & Export

The Replicability Principle

DashAI is designed so that any experiment can be reproduced. This is fundamental both for teaching (an instructor can repeat activities across cohorts) and for research (results must be verifiable).

What Gets Recorded

Every experiment in DashAI records:

  • Full configuration: task, dataset, splits, selected feature columns
  • Model parameters: all hyperparameters for each run
  • Evaluation metrics: per run, per split, and per level (step/batch/trial)
  • Data transformations: the converter chain applied to the dataset

This data is stored in SQLite (~/.DashAI/db.sqlite) with full history.

Traceability

Every action in DashAI leaves a record. This allows you to:

  • Compare runs with different configurations ("what did I change between run 1 and run 2?")
  • Audit methodological decisions
  • Communicate the process to third parties (reviewers, colleagues, supervisors)

The database schema separates ModelSession (experiment configuration) from Run (individual training execution), Metric (evaluation results), and Prediction (inference outputs).

Exporting Results

DashAI allows you to export results for use outside the platform:

  • Prediction results — downloadable from the Predictions view
  • Exploration plots — generated by Explorers within a Notebook and available for download as PNG images directly from the interface
  • Experiment configuration — visible in the UI for reference
warning

Full pipeline export (a portable "recipe" capturing the full preprocessing + model chain) is under development for future versions. The current version supports exporting results and basic configurations.

What is a Notebook?

A Notebook in DashAI is a working session with a mutable copy of a dataset. It groups Explorers (visualizations) and Converters (transformations) applied to that copy. The original dataset is never modified. See System Design → Notebook for details.

Data Storage Details

ArtifactStorage location
DatasetsApache Arrow IPC files in ~/.DashAI/
Trained modelsPickle/joblib files at ~/.DashAI/runs/{run_id}/
Optimization plotsSerialized Plotly objects alongside the run
MetricsMetric table in db.sqlite
Exploration resultsPNG images generated by Explorers within a Notebook, referenced by the Explorer table