rnow-rewards

What problem does it solve?

Reward function design for ReinforceNow RL training can be error-prone and time-consuming. This Skill provides a structured approach to implementing and validating reward functions, including common patterns and best practices.

Core Features & Use Cases

• Supports exact-match, contains, numerical tolerance, math-verify, llm_judge, and combined strategies.
• Provides precondition, sandbox, and LLM-based evaluation workflows for robust reward design.
• Real-world use: implement a reward to score a model's answer against a ground-truth key and gate improvements with preconditions.

Quick Start

Create rewards.py using @reward-decorated functions such as accuracy, and wire them into train.jsonl's rewards field. Install necessary dependencies, e.g., math-verify, and configure secrets if using llm_judge.