base_temperature

[hobbitlong]

base_temperature

[Ywandung-Lyou]

Should it be a sensitive hyper-parameter that should be fine-tuned carefully on the dataset?

[HobbitLong]

It just scales up the loss, you can ignore it or use 1 as you like.

[FraCorti]

Should it be a sensitive hyper-parameter that should be fine-tuned carefully on the dataset?

This is something that I'm also struggling to understand. Did you understand if it should be fine-tuned or not?

[KaiChen1998]

Should it be a sensitive hyper-parameter that should be fine-tuned carefully on the dataset?

This is something that I'm also struggling to understand. Did you understand if it should be fine-tuned or not?

@HobbitLong @zlapp
By dividing self.temperature first, and then times (self.temperature / self.base_temperature), it seems like self.temperature is useless and the actual temperature is always self.base_temperature, which is always 0.07, different from the default temperature 0.1 according to the README file?