Model Lab
Upload your data and Model Lab benchmarks 8 ML algorithms simultaneously — with automated ranking, diagnostics, SHAP feature importance, and a one-click research report. No ML expertise required.
Overview
8 ML algorithms
Random Forest, GBM, XGBoost, SVM, KNN, LDA, Naive Bayes, and Decision Tree — all compared in one run.
Automatic ranking
A composite score (Accuracy × 0.7 + CV × 0.3) ranks every model. The winner is highlighted automatically.
AI recommendation & report
Get a plain-language model recommendation plus a one-click HTML report with diagnostics and SHAP importance.
Leakage guardrails
Suspicious perfect scores and target leakage are flagged and excluded from the Best Model — so a high score you cannot trust never wins.
Tune, predict & simulate
Tune the winning model with Optuna, predict new data, compare models, and explore what-if scenarios — all from the saved model.
What you need
What you get
How it works
Results save automatically — no button to press
Every time you run an ML analysis in Statistical Lab and results appear on screen, the result is saved to Model Lab automatically in the background. You do not upload anything here — Model Lab accumulates results from Statistical Lab and ranks them.
Go to Statistical Lab and run an ML analysis
Open any ML analysis — Random Forest, GBM, XGBoost, SVM, KNN, LDA, Naive Bayes, or Decision Tree. Upload data, select variables, and run. The moment results appear, they are saved to Model Lab automatically.
Run more models on the same data
Go back and run a different algorithm on the same dataset. Each result saves automatically. Run as many as you want — Model Lab keeps them all.
| Task Type | Models in Statistical Lab |
|---|---|
| Classification | Random ForestGBMXGBoostSVMKNNLDANaive BayesDecision Tree |
| Regression | Random ForestGBMXGBoostSVRKNNDecision TreeRidgeLasso |
| Clustering | K-MeansK-MedoidsDBSCANHDBSCANGMMHCA |
Open Model Lab to compare
Open Model Lab from the dashboard. All saved results appear ranked by composite score. Each panel below shows what you will see.
Winner Summary
The top card shows the best-performing model with its Accuracy, CV Score, Generalization Gap, and a plain-language verdict on deployment readiness.
Model Rank Table
All models sorted by composite score. Columns show Accuracy, F1, AUC-ROC (classification), and CV Score ± standard deviation. The winning row is highlighted.
Auto Recommendation
Shows WHY THIS MODEL and WATCH OUT FOR based on your dataset — n, variable count, class balance.
Model Diagnostics
Each model is checked against four criteria — ✅ pass, ⚠️ warning, ❌ fail.
SHAP Feature Importance
Mean absolute SHAP values for the winning model. A longer bar = more influence. Features near zero can often be removed without affecting accuracy.
Deployment Readiness
A 5-point checklist: performance threshold, generalization gap, CV stability, sample adequacy, feature signal. Each is pass / warn / fail.
Click into any model to open its full detail card — metrics, variables, feature importance, confusion matrix, and model-health checks:
Export the report
Click Download Report at the top of the results panel to get a formatted HTML file — or use AI Chat to ask follow-up questions about the results.
Download Report
Click Download Report. An HTML file is generated with winner summary, rank table, diagnostics, SHAP, AI interpretation, and deployment readiness.
AI Chat
Open AI Chat to ask follow-up questions about the model ranking, feature importance, or next analytical steps. The AI has access to the current run's metrics.
Task types
Model Lab supports three task types. Each produces different results and uses a different set of metrics.
Classification
Auto-detectedPredicts which category an observation belongs to. Used when your target variable is categorical — e.g. spam/not-spam, species, customer segment.
Results shown
Diagnostics
Regression
Auto-detectedPredicts a continuous numeric value. Used when your target is a number — e.g. price, temperature, sales volume.
Results shown
Diagnostics
Clustering
No target variableGroups observations into natural clusters without a predefined label. Used for segmentation, pattern discovery, or anomaly detection. No target variable is required — only feature variables.
Results shown
Diagnostics
Guardrails — when a score is too good to trust
A perfect or near-perfect score is usually a warning, not a win. Model Lab automatically checks every saved model for target leakage, duplicated targets, and suspicious perfect scores. Flagged models are excluded from the Best Model, recommendations, and comparisons — so a score you cannot trust never rises to the top.
Example: R² = 1.000
If a model scores a perfect 1.000, it almost always means one of your feature columns secretly contains the answer (leakage). Model Lab marks it red, removes it from the leaderboard's top spot, and tells you to remove the offending column and re-run.
What gets flagged
What Model Lab does
Model card showing a leakage warning (Model Health)
red 'leakage suspected' banner on a flagged model
Tuning the winner
Once the quick comparison finds your top model, you can squeeze out more performance by tuning its hyperparameters. Model Lab uses Optuna to search within a time budget, then re-checks guardrails so a higher score never hides leakage. Only tree-based winners (XGBoost, Random Forest, GBM) can be tuned.
Run Auto Compare and find the winner
When the leaderboard settles and the top model is XGBoost, Random Forest, or GBM, a "Tune the winner" panel appears below it.
Pick a preset
Fast (~1 min), Balanced (~3 min), or Thorough (~4 min). Longer presets search more hyperparameter combinations and are more likely to find a better setting.
Read the result
Model Lab shows the before → after score gain, the best parameters found, the number of trials, and a fresh guardrail re-check verdict.
Use the tuned model
The tuned model is saved automatically — use it right away with Predict or the What-if simulator.
The trust difference: after tuning, guardrails run again. If the gain came from leakage, you get a red warning instead of false confidence — the higher number alone is never treated as success.
Tune the winner panel — presets and before/after gain
Fast / Balanced / Thorough buttons and the result card
Predict & compare
Every trained model is saved and reusable — it does not vanish when the analysis ends. Score new data with the saved model, or run the same data through several models to see where they agree and disagree.
Predict
Click Predict on any saved model card. Enter a single row, or upload a CSV for batch scoring. Download the predictions as a CSV.
Compare predictions
Use Compare predictions in the group header. Select two or more models, upload one CSV, and see each model's prediction side by side.
Predict dialog — single row and CSV batch
Compare predictions — models side by side
What-if simulator
Open What-if on any saved model to explore how its prediction responds as you change the inputs. Move a slider or pick a category and the prediction updates live.
Read it correctly: this shows how the model's predictionreacts to inputs — it is not the real-world effect of changing that variable. The model learned associations, not causes. Use it to understand the model, not to conclude “changing X causes Y.”
Start from the average, or load a row
Sliders and dropdowns start at the data’s average / most-common values. Or upload a CSV row to explore one specific case.
Change inputs and watch the prediction
Drag a slider or pick a category — the predicted value updates live as the model re-scores.
Sweep one variable
Pick a numeric variable and plot the model’s response curve across its whole range, holding the others fixed (like a PDP).
Needs a freshly saved model. The simulator reads variable ranges stored when a model is saved. Models saved before this feature was added show a short note — just re-run and save the model again to enable What-if.
What-if simulator — sliders, live prediction, and response curve