Consider using the RL algorithm -GRPO training？

[RyanOvO]

Consider using the RL algorithm -GRPO training？with SFT data.

[BradKML]

Seconding this, but with Dr. GRPO (de-biasing overthinking), Cross-Domain RL, Difficulty-aware RL, plus a few others for good measure.

[caoshiyi]

Hi @RyanOvO @BradKML, thanks for your interest! Please check out: https://novasky-ai.github.io/posts/sky-t1-7B/ and https://github.com/NovaSky-AI/SkyRL, some of our efforts in using RL to enhance the model's reasoning ability and agentic capability!

[BradKML]

@caoshiyi let's keep going with the conversations here, I have a lot of questions to ask NovaSky-AI/SkyRL#12