Reward functions

Reward functions turn generated completions or trajectories into numeric scores. They are task-specific Python files loaded by AReno’s training path and should be deterministic while you are debugging a run.

The public shape is:

def reward_fn(example, completions) -> list[float]:
    ...

The function receives the original example plus generated completions, then returns one score per completion. For agentic workflows, keep enough task state in the dataset row for the reward function to explain why a trajectory passed or failed.

Practical rules

  • Keep parsing and scoring explicit.

  • Return one score for each completion.

  • Avoid network calls in the hot path unless the task requires them.

  • Log enough context to debug wrong scores.

Where to go next