Easy Samples Are All You Need: Self-Evolving LLMs via Data-Efficient Reinforcement Learning

Create an open-source framework that implements 'Easy Samples' RL, focusing on data-efficient fine-tuning without massive manual labeling. This would help developers scale model training on limited high-quality data.

Suggested repo: easy-rl

"Reinforcement learning for LLMs, minus the reward hacking and high costs."

Estimated effort: 60h