diff --git a/_questions/machine-learning-zoomcamp/module-4/031_6862614ca5_combine-train-and-validation-before-test-eval.md b/_questions/machine-learning-zoomcamp/module-4/031_6862614ca5_combine-train-and-validation-before-test-eval.md
new file mode 100644
index 00000000..876f469a
--- /dev/null
+++ b/_questions/machine-learning-zoomcamp/module-4/031_6862614ca5_combine-train-and-validation-before-test-eval.md
@@ -0,0 +1,17 @@
+---
+id: 6862614ca5
+question: Why do we combine the training and validation datasets before evaluating
+  the final model on the test dataset?
+sort_order: 31
+---
+
+The validation set is used during model selection and hyperparameter tuning (i.e., to estimate performance and decide which model configuration to keep).
+
+Once you’ve selected the final model, you no longer need the validation set for tuning—so it’s common to retrain the final model on the combined data (train + validation) to let it learn from as much labeled data as possible.
+
+You should still keep the test set untouched until the very end, because the test set is reserved for the final, unbiased evaluation on unseen data.
+
+Example:
+- `df_train_full = pd.concat([df_train, df_val])`
+- Train the final model using `df_train_full`
+- Report performance only once on the held-out test dataset
\ No newline at end of file