Details about the DPO training config #9
Description
Thank you for your excellent work on the recent paper. I noticed that some training hyperparameters were provided, such as:
- Learning rate: 1e-4
- Iterations: 1500
- Number of data pairs: 6,000
I’m writing to ask if you could kindly share more details about the DPO training setup—specifically the batch size, gradient accumulation steps, and the value of beta_dpo used during training.
Thank you in advance for your time and assistance.
Activity
Ephemeral182 commentedon Jul 6, 2025
Thank you for your interest in our work. For the DPO training setup, we used a total batch size of 16, and set beta_dpo to approximately 5000–10000.