Unlike other repositories that focus more on curve analysis, we emphasize interesting phenomena and process analysis to make it easier for everyone to understand what is happening.
This repository's code is inspired by DeepSeek-RL-Qwen-0.5B-GRPO-gsm8k and deepseek_r1_train and is not original. I have only conducted experimental analysis based on it. Thanks for their endeavor!
Don't use models that are not INSTRUCTION! But, Why?
Qwen-2.5-3B output
Qwen-2.5-3B-GRPO output