Track Experiments

MLOps Desktop includes full experiment tracking: run history, experiment grouping, model registry with versioning, and side-by-side comparison. All data stays local in SQLite.

Time to complete: ~10 minutes

Prerequisites

• Run several pipelines with different configurations
• Familiarity with the Quickstart

Runs Tab

The Runs tab shows all pipeline executions with metrics.

View Modes

Flat View

All runs in a single list, sorted by date (newest first).

Column	Description
Name	Run display name (editable)
Pipeline	Pipeline that was executed
Experiment	Associated experiment (if any)
Metrics	Key metric (Accuracy or R²)
Duration	Execution time
Status	Success/Failed

Hierarchical View

Runs grouped by experiment.

▼ Titanic Classification (3 runs)
    Run 1: Accuracy 0.82
    Run 2: Accuracy 0.85
    Run 3: Accuracy 0.87
▼ Iris Baseline (2 runs)
    Run 1: Accuracy 0.96
    Run 2: Accuracy 0.97
▼ Unassigned (5 runs)
    ...

Run Actions

Click a run to:

• View metrics — Full metrics display in Metrics tab
• View explain data — Load SHAP/importance from this run
• Edit metadata — Change display name, add notes, tags
• Register model — Add to model registry
• Delete — Remove run (with confirmation)

Experiments

Experiments group related runs for organization.

Create an Experiment

1. Click the Experiment selector dropdown in the toolbar
2. Click + New Experiment
3. Enter a name and description
4. Click Create

New runs will automatically be associated with the active experiment.

Experiment Status

Status	Meaning
Active	Currently running experiments
Completed	Finished experiments (archive candidates)
Archived	Hidden from default view

Filter runs by experiment status in the Runs tab.

Compare Runs

Compare 2-5 runs side-by-side:

1. In the Runs tab, select multiple runs (Cmd+click)
2. Click Compare
3. View the comparison modal

Comparison View

                    Run 1      Run 2      Run 3
Accuracy            0.82       0.85       0.87 ✓
Precision           0.79       0.83       0.85 ✓
Recall              0.78       0.81       0.84 ✓
F1 Score            0.78       0.82       0.84 ✓

Parameters:
n_estimators        100        150        200
max_depth           10         15         15
learning_rate       -          -          0.1

The best value in each row is highlighted.

Model Registry

The Models tab manages your model versions.

Register a Model

1. In the Runs tab, click a successful run
2. Click Register Model
3. Enter model name (or select existing)
4. Add description, tags
5. Click Register

Model Stages

Each model version has a stage:

Stage	Purpose	Typical Use
None	Just registered	New models
Staging	Testing/validation	Pre-production
Production	Serving live traffic	Current best model
Archived	No longer active	Old versions

Promote models through stages:

None → Staging → Production → Archived

Model Actions

• View versions — Expand to see all versions
• Compare versions — Side-by-side metrics comparison
• Promote/demote — Change stage
• View details — File path, format, metrics snapshot
• Delete — Remove version (with confirmation)

Tagging

Add tags to runs and model versions for organization:

Tags: baseline, tuned, production-candidate, best-so-far

Filter by tags in both Runs and Models tabs.

Tag Best Practices

Tag	Use For
baseline	Reference model to beat
tuned	After hyperparameter optimization
production	Currently deployed
experiment-v2	Group related experiments
best	Your top performer

Annotations

Add notes to runs for context:

Display Name: "RF with feature engineering"
Notes: |
  - Added age_group feature
  - Removed cabin column (too many nulls)
  - Stratified by Survived
Tags: [feature-engineering, v2]

Access via Edit Metadata on any run.

Best Practices

Always Use Experiments

Group related runs into experiments:

Experiment: "Titanic - Model Selection"
  Run 1: Logistic Regression
  Run 2: Random Forest
  Run 3: Gradient Boosting

Experiment: "Titanic - RF Tuning"
  Run 1: n_estimators=100
  Run 2: n_estimators=200
  Run 3: Optuna 50 trials

Tag Important Runs

When you get a good result, tag it immediately. It’s easy to forget which run was best.

Compare Before Deploying

Before promoting a model:

1. Compare against current production
2. Verify improvement on all key metrics
3. Check explain data for reasonable feature importance

Clean Up Failed Runs

Delete failed or test runs to keep history clean. Use filters to find them:

• Status: Failed
• Duration: < 5s (likely errors)

Data Storage

All tracking data is stored locally:

~/Library/Application Support/com.mlops.desktop/settings.db

Tables:

• runs — Run metadata, status, duration
• run_metrics — Metrics per run (JSON for complex data)
• models — Model registry
• model_versions — Versions with stage, tags
• experiments — Experiment metadata
• run_annotations — Display names, notes, tags

Tip

All data stays on your machine. Nothing is uploaded to the cloud.

Export Data

Export run history for external analysis:

import sqlite3
import pandas as pd

conn = sqlite3.connect(
    "~/Library/Application Support/com.mlops.desktop/settings.db"
)
runs = pd.read_sql("SELECT * FROM runs", conn)
metrics = pd.read_sql("SELECT * FROM run_metrics", conn)

Next Steps

Deploy Your Model

Serve models via HTTP API

Model Export Formats

joblib, pickle, and ONNX options