reinforcement learning improvement

[Ski-ing]

How significant is the improvement in code generation performance metrics attributed to the Group Relative Policy Optimization (GRPO) within the reinforcement learning component?

[DeepSeekPH]

The performance of GRPO varies depending on the test sets. Generally, GRPO demonstrates an improvement of approximately 0.5 points on code generation test sets. The enhancements on math-related benchmarks are more substantial.

[Ski-ing]

Thanks for your reply