Tabular machine learning · from data to deployment
InferLabs takes your dataset from profiling and validation to model training, evaluation, versioning and prediction, in one guided workflow.
Workflow
The same steps, always in the same order, so every model you keep can be explained and reproduced.
Bring your CSV into InferLabs.
Profile the data, find quality issues, read the target.
Preprocessing compared by cross-validation, per model.
Several models, scored fairly on the same folds.
A controlled hyperparameter search for supported models.
Diagnostics and generalization on held-out rows.
Reproducible versions with their configuration.
The saved pipeline makes the predictions.
Understand
InferLabs doesn't train blindly. It checks what is inside the dataset first: missing values, duplicates, columns that look like IDs, and what the target actually represents.
It reads the content of the target, not just its data type: numbers stored as text, categories, discrete values and identifiers are each explained.
Compare
Every model is scored with cross-validation on the training rows, using the preprocessing that suits it. The recommendation comes from those scores; the held-out test rows are only used for the final check.
Which model wins depends on your data. The values shown are illustrative.
Evaluate
Each model is graded against a mean-value baseline and fold-to-fold noise, then checked on rows it has never seen, with diagnostics that show where it is wrong.
A weak model is never called good.
Version
A version stores the model together with its exact preprocessing and hyperparameters. An optimized version is kept only when cross-validation improves; nothing is activated behind your back.
Missing values are imputed inside each fold. Nothing is removed automatically.
model_paramsRegression and classification
Predict a number.
R² · RMSE · MAE
Predict a category.
Accuracy · F1 · ROC-AUC
Numerical featuresCategorical featuresMixed datasets work for both: only the target decides the task.
Capabilities
No leakage between folds
Imputation, encoding and scaling are learned from each fold's training rows only. Validation and test rows are transformed, never fitted.
Dataset health
Reported and explained. Nothing is removed for you.
Class imbalance illustrative
3.9 : 1 ranked by F1, class weights on
Accuracy can mislead on imbalanced labels, so F1 and ROC-AUC rank the models.
Hyperparameter optimization illustrative
+0.030 CV R² over the baseline
CV R². The tuned version is kept only when cross-validation improves beyond fold noise.
Metrics that fit the task
R² · RMSE · MAE, compared with a mean-value baseline
Accuracy · balanced accuracy · precision · recall · F1 · ROC-AUC
Prediction illustrative
The row goes through the version's saved preprocessing and model. Nothing is retrained.
Know what is inside your dataset before training.
Several models, the same cross-validation.
A controlled search, kept only if CV improves.
Artifacts and configuration, reproducible.
Upload a dataset, understand what it contains, compare models, and move from experiment to prediction.
Contact
InferLabs is built and maintained by Saurav Neupane. Reach out for feedback, collaboration or a walkthrough.
Version 1
Restoring your session…