intfloat

/

multilingual-e5-large

like 1.19k

Hello, great model. I tested on English, but on my language (Kazakh it's not working well). What would be the best way to fine-tune it? I have prepared dataset 500k rows of wiki data (1 row 512 max token chunks) of (anchors, positive, negative) on 3 languages (ru, kk, en) and want to fine-tune this model, but I'm not sure what would be the right way to do it. For example:

anchor (en) | positive (en) | negative (en)
or
anchor (en) | positive (kk) | negative (kk)
or
anchor (kk) | positive (en) | negative (en)

or combined all cases as a 1 dataset.

Don't have intuition how exactly do it, please if someone can help that will really help us 🙏.