forward-looking-examples

What problem does it solve?

Producing high-quality forward-looking forecasting datasets is hard because questions must be specific, verifiable, temporally resolvable, and formatted consistently for GRPO-style training.

Core Features & Use Cases

• End-to-end dataset pipeline templates (GRPO): seeds → forward-looking question generation → optional context → linting → temporal filtering/splitting → training.
• Domain-specific examples: ready-to-adapt setups for golf outcomes, Trump policy actions, military strike events, and general forecasting using GDELT, plus timestamped-document RAG workflows.
• Quality control and leakage prevention: lint results are used to remove affected samples, and temporal splitting + horizon constraints reduce answer leakage.

Quick Start

Ask your assistant to generate a GRPO-ready forecasting dataset using NewsSeedGenerator (or GdeltSeedGenerator/FileSetSeedGenerator) with ForwardLookingQuestionGenerator, then lint the full dataset, filter to the desired resolution horizon, and run training with GRPOTrainingConfig.