was looking to finetune an adapter as

I think its show the ada is better than bge, but bge can be slightly better with fine-tuning (for this specific dataset)

(tbh I just quickly skimmed it lol)

what metrics show that bge can be slightly better in this article? both hit rate and mrr is lower. (0.803797 vs 0.87088, 0.667426 vs 0.728840 at best)

don't understand the point of this article if it yields poorer metrics

or am i misunderstanding something?

bge has hit rate of 0.78, then with the fine-tuned adaptor its 0.79

right

so small increase when finetuning an adaptor for bge

but i was under the impression that it would be better than ada

is it supposed to be better if we fine-tune with some other text embedding models?

What gave you that impression? Especially since it's bge-small, its quite impressive performance compared to ada imo

People want local embeddings, and they want them to be good. This is one option to improve performance of local embeddings 🤔

i see. i was under the impression that this was supposed to tell us how to fine-tune an embedding model that's better than ada.

thank you for your answer.