Forward-Looking Training Examples (GRPO)

What it does

Provides production-ready example pipelines for generating forward-looking binary forecasting datasets (GRPO). The skill documents end-to-end patterns: seed generation (news, GDELT, or filesets), question generation with ForwardLookingQuestionGenerator, context generation, automated labeling, linting, temporal splitting, and training configuration recommendations for fine-tuning models. Multiple domain examples (sports, politics, military, general forecasting, timestamped documents) illustrate practical parameters and notebooks to reproduce results.

When to use it

Use when building forecasting datasets or experimenting with GRPO-style fine-tuning. Useful for teams preparing high-quality labeled forecasting data, testing question pipelines on small seeds, or scaling to production datasets (thousands of seeds). Not intended for casual Q&A — it's for dataset engineers and modelers.

What's included

• Scripts: none declared under the skill dir, but many in-repo notebooks and example pipelines referenced (notebooks for golf, Trump policy, military strikes, GDELT, FileSet RAG).
• References: example notebooks and notebook paths referenced in the skill body and repo.
• Instructions: concrete code examples showing pipeline construction, training/config recommendations (gpt-oss-120b, lora_rank, num_rollouts), dataset linting and temporal split patterns, and domain-specific guidance (intervals, questions_per_seed, answer formatting).

Compatible agents

Python SDK consumers and engineers using LightningRod tooling; compatible with agents that can run Python SDK calls, notebook-driven workflows, and training orchestrations (local Python agents, Jupyter/Colab, and CI-driven training pipelines).