Comparision with Marqo Siglip

[aretius]

Hello
Thanks for sharing the e-commerce embedding model thats beating SOTA by a nice margin.
I wanted to understand how are the e-commerce embeddings different from Marqo Siglip trained earlier from a metric standpoint.
Did you ever do a comparison on same data for retrieval/search?

[jn2clark]

Hi @aretius , there are a few differences. The main differences were the data sets. FashionSigLIP was trained with a smaller but richer dataset. This also changed how the loss was done between the two models. We used 7 text fields and 1 image field so the loss was all combinations of img and text as well as the mean of the text vectors (i.e. fused).

[aretius]

Got it so the dataset and loss was different.
Did you also evaluate both models on text/image based retrieval? If yes what difference did you notice?