HIT-TMG

/

KaLM-embedding-multilingual-mini-instruct-v1.5

like 55

Follow

HITsz-Text Machine Group 46

When are you planning to release this model?

Hi, Gurur

Thank you for your attention and support.

Due to the large number of parameters, the optimization progress of KaLM-embedding-multilingual-max-instruct-v is currently slow.
The specific release date is not yet determined, perhaps within one or two months.

Hi, YanshekWoo

Thanks for information. I'm happy with performance of HIT-TMG/KaLM-embedding-multilingual-mini-instruct-v1.5

Also would like to see reranker from you.

RE: Pooling (mean vs last token)

I was just interested in it as most frameworks (TEI/TRT) will perform variable length inference. For pooling=mean, there is no efficient variable length mean pooling. Such a kernel would be easy to write, but hard to get it right without sacrificing precision.

Its also one of the reasons modernbert has cls/eos pooling.
https://huggingface.co/Alibaba-NLP/gte-modernbert-base/blob/main/1_Pooling/config.json

It would be cool to see the max trained with last token pooling

RE: Pooling (mean vs last token)

I was just interested in it as most frameworks (TEI/TRT) will perform variable length inference. For pooling=mean, there is no efficient variable length mean pooling. Such a kernel would be easy to write, but hard to get it right without sacrificing precision.

Its also one of the reasons modernbert has cls/eos pooling.
https://huggingface.co/Alibaba-NLP/gte-modernbert-base/blob/main/1_Pooling/config.json

It would be cool to see the max trained with last token pooling

Thank you for your feedback. Taking these factors into consideration, we will prioritize replacing the pooling method with last token pooling.