RL environments, SFT datasets, DPO datasets, prompt packs, and other artefacts that lift models on specific evals.