forward-looking-examples

Generate a GRPO-ready forecasting dataset by combining seed generation with forward-looking question generation, optional news context, labeling, linting, and temporal filtering. This pipeline produces specific, verifiable, and temporally resolvable questions formatted consistently for GRPO-style training.

Prevent answer leakage in forecasting datasets by applying temporal splitting and horizon constraints during the filter_and_split phase. Linting results are also used to remove affected samples, ensuring the remaining data is clean and temporally isolated for training.

Yes, you can use GdeltSeedGenerator to create training sets for predicting military strike events. The pipeline supports domain-specific examples including military strikes, politics, and sports by leveraging GDELT or timestamped FileSet RAG for context.

Temporal splitting works by applying days_to_resolution_range constraints during the filter_and_split phase. This filters questions based on their resolution horizon, reducing answer leakage and ensuring the dataset is properly segmented for valid forward-looking training.

Yes, the Lightning Rod SDK is required to lint and split forecasting datasets. You must use SDK components like NewsSeedGenerator and ForwardLookingQuestionGenerator, then execute lr.datasets.linter.run followed by filter_and_split to finalize the data.

You can generate forward-looking questions for GRPO training spanning sports outcomes, political policy actions, military strike events, and general forecasting. The pipeline ensures questions are specific, verifiable, and temporally resolvable using news, GDELT, or timestamped documents.