The final test-set labels.

The number of reinforcement-learning rewards.

The exact names of all hidden clusters.

The number of clusters K.

Choose the model with the most parameters without validation.

Report only the best training score for each model.

Use a different test set for each model to make results more diverse.

Evaluate all models on the same split or cross-validation protocol and the same metrics.

The number of clusters in K-means before training.

The expected value of taking an action in a state and following a policy afterward.

The variance preserved by principal components.

The exact text frequency of all tokens in a corpus only.

Finding density-based clusters and identifying noise points.

Creating labels for supervised learning without any data.

Replacing feature engineering in every problem.

Fitting a straight line to predict a continuous target only.

To estimate model performance more reliably by using multiple train/validation splits.

To make all features categorical.

To remove the need for a test set in every project.

To hide evaluation results from the researcher.

A sequence of rewards without features.

A continuous numerical value.

Only image segmentation masks.

Only unordered clusters.

Use one decision tree as the whole reinforcement-learning environment.

Use a neural network to approximate Q-values.

Use TF-IDF to forecast stock prices without time order.

Use PCA to remove target labels.

Increasing dataset size without optimization.

Using batching, parallel preprocessing, multi-core training, or GPU acceleration when appropriate.

Running the same slow code repeatedly without measurement.

Removing all documentation from the project.

Train models sequentially so later models focus more on previous errors.

Replace all features with missing values.

Use no evaluation metrics.

Train only one model and never update it.

Combine many decision trees to improve stability and generalization.

Use one linear equation for every possible task.

Use only one feature and ignore all others.

Cluster data without any trees.