Hi Zindi team,

A rules clarification on using the test set during training.

Is it permitted to use the provided test images in training via self-training / transductive learning, specifically:

1. run our own trained models on the test images,

2. keep only high-confidence lines where several models agree (labels produced entirely in code, never manually),

3. add those test images + their model-generated labels back into the training set and continue training?

The constraints on our side:

- No external data; only the images provided for this challenge are used.

- No external labels; the pseudo-labels are generated in code by our own models from the provided test images; nothing comes from outside the competition.

- No manual labeling.

I ask because the rules say "you may use only the datasets provided" (the test set is provided data), but the clarification in the "hyperparameter search & training-data cleaning" thread noted decisions should be made "with no use of test data or external labels." Could you confirm whether that applies only to training-data cleaning, or whether using the test images for self-training is also disallowed?

Thanks!