Integrate GRPO for Replicating DeepSeek-R1 and Other Inference Models, Achieving a HuggingFace open‑r1–like Effect

[submartingales]

⚠️ Please check that this feature request hasn't been suggested before.

• I searched previous Ideas in Discussions didn't find any similar feature requests.
• I searched previous Issues didn't find any similar feature requests.

🔖 Feature description

We need to support replicating the performance of DeepSeek‑R1 on our existing model. The main requirement is to integrate the GRPO algorithm, which has already been implemented in git+https://github.com/huggingface/trl.git.

✔️ Solution

We need to support replicating the performance of DeepSeek‑R1 on our existing model. The main requirement is to integrate the GRPO algorithm, which has already been implemented in git+https://github.com/huggingface/trl.git.

❓ Alternatives

There have been ongoing efforts to support reproducing DeepSeek-R1 on https://github.com/huggingface/open-r1

📝 Additional Context

The GRPO algorithm has already been implemented in git+https://github.com/huggingface/trl.git.

Acknowledgements

• My issue title is concise, descriptive, and in title casing.
• I have searched the existing issues to make sure this feature has not been requested yet.
• I have provided enough information for the maintainers to understand and evaluate this request.