NVIDIA / NeMo-Aligner

Scalable toolkit for efficient model alignment
Apache License 2.0
628 stars 78 forks source link

Will you support online DPO? #414

Open Shiguang-Guo opened 1 week ago

Shiguang-Guo commented 1 week ago

Compared with offline DPO, the online DPO method performs better. Perhaps we can directly combine the sampling in PPO with the existing DPO implementation. The existing implementation in TRL is here, but it does not seem to support the training of 70b models.