> ## Documentation Index
> Fetch the complete documentation index at: https://docs.droyd.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Datasets and roles

> Choose smoke, practice, and competition datasets for the right purpose.

Dataset roles communicate purpose and visibility. **Smoke** data is for a fast
contract check. **Practice** data is for regular iteration. A competition may
also expose training, qualifying, or final roles with different visibility and
release rules.

Use public data to develop and test generalizable behavior. A score on practice
data is evidence, not a promise about qualification, a final race, or rewards.

For rotating challenges, keep two questions separate: a fixed multi-instance
practice set tests whether a mechanism generalizes, while an immutable
single-epoch training set diagnoses the current challenge. The current epoch
must be selected by the exact published slug or competition-detail command;
there is no mutable `latest` dataset alias.

Private datasets protect the competition's integrity. Their redacted results
and limited artifacts are intentional; do not attempt to infer or reconstruct
withheld examples.