Will the different inputs of different training steps affect the model effect?

[vincezengqiang]

hihi thanks for the excellent work!
I found that the sftdataset will rewrite by prompt_dict，and the reward model is trained by raw data， and in input of inference.py just like：'Question: How are you ? Answer:'。
so i m curious that will difference inputs of different training steps affect the model effect？
Looking forward to your reply, thanks!

[binmakeswell]

Hi @vincezengqiang Sure, dataset is crucial to the model performance. This is the reason why some people say "models are data-centric"