Dataset loader API

Dataset loaders normalize external records for the selected training path. The exact record shape depends on the algorithm family.

Current loader shapes are documented in Dataset loaders:

• SFT loaders provide prompts and target responses.

• DPO loaders provide prompt, chosen, and rejected responses.

• Prompt-based RL loaders provide prompts and task metadata for reward scoring.

• Agentic RL loaders provide task state for agent functions.

Keep loaders small, deterministic, and testable outside the GPU training loop.