examples-guide

To choose a training pattern, use a decision tree based on your dataset and prediction goal: apply forward-looking GRPO for forecasting, content-learning SFT for teaching domain facts, or tabular processing for mapping structured data.

The best answer type framing for forecasting tasks depends on your signal quality needs: use binary or multiple choice for reliable labeling, numeric for continuous predictions, or free response to capture open-ended outcomes.

To avoid label leakage during temporal splitting, lint your dataset before splitting and ensure that all prediction dates strictly precede their corresponding outcome dates to maintain forecasting validity.

You should use tabular sample mapping instead of content-learning SFT when your workflow involves processing structured tabular data rather than teaching a model general domain facts through text.

Key guardrails for building a forward-looking GRPO dataset include enforcing temporal splitting, preventing label leakage, linting data before splits, and verifying that prediction dates precede outcomes.