Training and evaluation of the slime Search-R1-trained Qwen2.5-3B model on NQ and HotpotQA.

Comparing PPO and GRPO for post-training Qwen2.5-0.5B-Instruct on GSM8K with veRL, focusing on a small-model, verifiable-reward setting where the reward is rule-based correctness rather than a learned reward model.

From CartPole-v1 to CartPole Swing-Up