rnow-rewards

To define reinforcement learning rewards, you create a rewards.py file using @reward-decorated functions and wire them into train.jsonl's rewards field. This provides a structured approach to implementing and validating reward patterns for ReinforceNow training.

Available reward patterns for validating model outputs include exact matches, contain-based checks, numerical tolerance, math verification, and LLM-based judgments. You can also combine these strategies to robustly evaluate training responses.

Yes, you can use an LLM judge to score reinforcement learning responses by configuring the llm_judge reward strategy. This requires an OpenAI API key configured as a secret to evaluate and score the model's answer.

You need to install the math-verify dependency to implement numerical tolerance rewards and math verification workflows. This optional tool provides robust mathematical evaluation when scoring model outputs against ground-truth keys.

To gate reinforcement learning improvements with preconditions, you use the skill's precondition workflows to evaluate a model's answer before applying the reward. This ensures only valid responses pass the gate before scoring against ground-truth keys.

The best way to structure reward functions for ReinforceNow is using @reward-decorated functions like accuracy in a rewards.py file. This pattern supports sandbox execution and combined strategies for robust reward design.