Gen-Verse/CURE

Uses a co-evolutionary RL framework where LLM coders and unit testers improve simultaneously without requiring ground-truth test cases—the tester learns from coder failures during rollouts. Provides modular training pipeline (sampling, execution, reward assignment, training) supporting vLLM and API-based inference, with comprehensive evaluation across one-shot coding, test generation, and Best-of-N scaling scenarios. Includes pre-trained ReasonFlux-Coder models (4B/7B/14B) and six public datasets (CodeContests, LiveBench, MBPP, etc.) for reproducible code-generation RL research.

159 stars. No commits in the last 6 months.