Data requirements for DPO fine-tuning

[22TonyFStark]

Hi,
I'm using DPO for fine-tuning and would like to know: how much human preference data (i.e., preference pairs) is needed?
Any guidance on typical/minimum recommended amounts, or how it scales with model size?
Thanks!
Zhongfei Qing

[LukeLIN-web]

I heard 1k clips could work