intfloat

/

e5-mistral-7b-instruct

like 386

We will not release "official" fine-tuning script. Instead, we recommend using off-the-shelf embedding fine-tuning libraries such as Tevatron. You only need to change the tokenization, and pooling part.

@intfloat it looks like the pooling part in Tevatron is the same as in the paper (using hidden state of eos token). Therefore, only tokenization needs to be changed right?

@Mengyao00 @serialcoder
Model finetuning using huggingface peft + deepspeed
https://github.com/kamalkraj/e5-mistral-7b-instruct/

@intfloat

@serialcoder
In which files can I find the pooling and tokenization parts of Tevatron?
I can't find them.