Better questions.
Measured answers.
Choose your inputs. Validate your model. Predict on new data.
A
local workspace for practical machine learning.
Evidence before action
Selected model · reserved final test
Final test performance
Which inputs mattered?
Final test permutation importance: score lost when shuffled (three repeats). Describes model dependence, not causation. For loss metrics, a positive value means the error increased.
Reproducibility & split audit
Try a prediction
Alternate models are exploratory. Final test metrics and feature importance above belong to the selected model only. Classification probabilities are uncalibrated estimates.
For single and batch predictions only. Lower thresholds flag more positives. The final test report above uses the original estimator decision rule; changing this threshold does not revalidate it.
Predict a whole CSV
Include every selected input column with the same names. Extra columns are preserved. If the target is included, observed performance is calculated. Batch checks compare incoming data with development data.
Experiment history
THIS BROWSERLast 20 summaries stored on this browser, without datasets or model files. Models must be saved separately. Browser storage can be cleared; download reports for a durable record.
Working guide
Choose a validation strategy
Random: rows represent independent examples. Grouped: keep each customer, device, or location together. Time: train on earlier dates and test on later dates. Time validation is tabular prediction; it does not automatically create lags or a forecasting model.
Make the comparison meaningful
Balanced accuracy gives each class equal weight; macro F1 balances precision and recall across classes. MAE is average absolute error in target units; RMSE penalizes large errors more. R² compares against predicting the test mean and can be negative. Fold standard deviation measures variation, not a confidence interval.
Know where the boundary is
This static tool supports tabular experiments and manual batch scoring. It does not provide shared accounts, scheduled jobs, an inference API, live production monitoring, or domain-specific approval. Validate results with representative external data and a person who understands the business.
Data and model privacy
Data and model files stay in this browser unless you save or share them. Runtime packages download from external CDNs. A project file contains your dataset. A saved model can retain training examples (especially nearest neighbors); treat it as sensitive. Only load model files you trust. History summaries include names and metric results.
Dataset limits: 8 MB, 20,000 rows, 100 columns. Model imports: 40 MB. Larger data may exceed browser memory. Cancel resets the ML worker and clears in-memory models.