Run every model.
Find the best one automatically.
Upload your data and Model Lab benchmarks 8 ML algorithms simultaneously — with leakage guardrails, hyperparameter tuning, prediction, what-if simulation, and a one-click research report.
01 · MODEL COMPARISON
Run 8 models at once. Pick the winner.
One dataset. Eight algorithms. Automatic ranking.
Upload your data and Model Lab runs all applicable models simultaneously — classification or regression. A composite score ranks every model, highlighting the best performer without manual trial and error.
- 8 ML algorithms — RF, GBM, XGBoost, SVM, KNN, LDA, NB, DT
- Composite score ranking — objective, reproducible
- Side-by-side metrics: Accuracy, F1, AUC, CV, Gap
- Auto-detects classification vs regression, auto-selects variables
| # | MODEL | ACCURACY | F1 | CV |
|---|---|---|---|---|
| 🥇 | LDA | 100.0% | 1.000 | 98.0±2.7% |
| 🥈 | Random Forest | 97.3% | 0.973 | 96.1±3.1% |
| 🥉 | SVM | 96.7% | 0.967 | 95.3±4.5% |
| 4 | SVM | 96.7% | 0.967 | 95.3±4.5% |
| 5 | Decision Tree | 93.3% | 0.933 | 95.3±3.4% |
| 6 | Random Forest | 90.0% | 0.900 | 94.7±2.7% |
02 · GUARDRAILS
A 100% score is usually a problem, not a win.
We check whether the score is real — not just high.
Most tools celebrate a perfect score. Model Lab does the opposite: it flags target leakage, duplicated targets, and suspicious perfect scores, then excludes those models from the leaderboard and recommendations. The trust layer no chatbot gives you.
- Detects target leakage, duplicated targets, suspicious perfect scores
- Leaky models excluded from Best Model, recommendations, and comparisons
- Class imbalance and overfit-gap warnings, severity-tagged
- Re-checked even after tuning — a higher score never hides leakage
Leakage suspected — excluded from Best Model
03 · DIAGNOSTICS & READINESS
Overfitting? CV instability? Ready to ship?
Every model checked, then a clear go / no-go verdict.
Model Lab evaluates overfitting risk, CV stability, sample adequacy, and predictive performance — then rolls them into a deployment-readiness verdict. Green, amber, or red. No ambiguity.
- Overfitting check: train-test gap threshold by category
- CV stability: fold-level variance flagged if σ > 5%
- Sample-size adequacy: dynamic threshold per variable count
- Deployment verdict: Ready / Needs Review / Not Ready, with what to fix
Linear Discriminant Analysis (LDA)
With n=150 and 4 numeric features, LDA handles linear boundaries exceptionally well.
WHY THIS MODEL
WATCH OUT FOR
04 · ALGORITHM ADVISOR
Which model fits your data — and what to analyze next.
Context-aware suggestions, not generic advice.
Model Lab reads your dataset — sample size, variable count, class balance — and recommends the right algorithm. After you find the winner, it points you to the next analysis (feature importance, decision tree, segmentation) and jumps you there in one click.
- Reads n, variables, class balance, data structure
- Plain-language justification for every recommendation
- Best model → next analysis, with variables pre-filled
- Flags edge cases: imbalanced classes, small n, high dimensionality
Mean |SHAP| values · n=150
05 · FEATURE IMPORTANCE & COMPARISON
Which variable actually moves the needle?
SHAP values, permutation importance — and cross-model agreement.
See model-native importance and SHAP values for every predictor. Then compare importance across all models in one table — variables that every model agrees on are the signals you can trust.
- SHAP mean |φ| — model-agnostic, consistent across algorithms
- Permutation importance — direct effect on accuracy
- Cross-model importance table — "robust" badge for agreed-upon variables
- Sortable — find disagreements between models fast
06 · HYPERPARAMETER TUNING
Squeeze more out of your winning model.
Optuna search within a time budget — then re-checked for leakage.
Pick a preset — Fast, Balanced, or Thorough — and Model Lab tunes the winning model with Optuna, searching hyperparameters within a time budget. It shows the before/after gain, the best parameters, and re-runs guardrails so a higher score never hides leakage.
- Presets map to time budgets — Fast (~1m) / Balanced (~3m) / Thorough (~4m)
- Bayesian (TPE) search via Optuna — smarter than grid/random
- Before → after gain, best parameters, trials run
- Post-tuning guardrail re-check — the gain is real, not leakage
Thorough preset · 120 trials · re-checked clean
07 · PREDICT & SIMULATE
Your model doesn’t vanish when the analysis ends.
Save the model, predict new data, and explore what-ifs.
Every trained model is saved and reusable. Score new data one row at a time or in bulk, compare how different models predict the same rows, and explore how predictions respond as you change inputs — clearly labeled as model response, not causation.
- Trained models persisted — reuse without re-training
- Predict new data: single row or CSV batch, downloadable
- Compare predictions across models — see where they disagree
- What-if simulator: watch predictions react (not a causal claim)
Model response to inputs — not a causal claim
Winner Summary
🏆 Linear Discriminant Analysis (LDA)
Composite Score: 99.4 · Accuracy: 100.0% · CV: 98.0%±2.7%
AI Interpretation
The LDA model achieved excellent performance (Accuracy=100.0%, CV=98.0%±2.7%) with strong generalization (Gap=2.0%). Petal.Length was the most influential feature (0.5 SHAP)...
08 · AUTO REPORT
One click. A research-grade report.
Download a complete model comparison report.
Model Lab generates a structured report covering winner summary, full rank table, diagnostic results, guardrail checks, feature importance, and an AI-written interpretation.
- Winner summary, rank table, diagnostics, guardrails, importance
- AI-written model interpretation
- One-click download — print to PDF from browser
- Clean, professional layout — no extra formatting needed