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.