You should not use test samples to generate training data. The core principle is that the test set is meant to evaluate your model's ability to generalize to unseen data. Using test data during training or data augmentation violates this principle and can lead to overfitting.

While the specific rules for the "Buzz: Data augmentation for phase 2" challenge should always be the final authority, Zindi's standard rules emphasize:

• Submissions must run on the original, provided datasets.
• Using data in ways that create leaks (like using the test set for training) is against the spirit of fair competition and can lead to disqualification, as it produces solutions that are not valuable to the client.