NVlabs/GDPO
GDPO is a reinforcement learning training method for improving LLMs on math reasoning, code generation, and tool-calling tasks.
Not currently ranked — collecting fresh signals.
star history
GDPO addresses reward advantage collapse in Group Relative Policy Optimization (GRPO) when handling multiple rewards. It decouples reward normalization across individual rewards to preserve relative differences and enable more stable LLM fine-tuning. The implementation integrates with major RL training frameworks including TRL, VERL, and NeMo-RL, providing SLURM-free training scripts that can run on 8xA100 GPUs in approximately 1 hour.
Frequently asked
- What is NVlabs/GDPO?
- GDPO is a reinforcement learning training method for improving LLMs on math reasoning, code generation, and tool-calling tasks.
- Is GDPO open source?
- Yes — NVlabs/GDPO is open source, released under the Apache-2.0 license.
- What language is GDPO written in?
- NVlabs/GDPO is primarily written in Python.
- How popular is GDPO?
- NVlabs/GDPO has 480 stars on GitHub.
- Where can I find GDPO?
- NVlabs/GDPO is on GitHub at https://github.com/NVlabs/GDPO.