model_card_data not available on CrossEncoder

[x-tabdeveloping]

CrossEncoder objects currently do not have the model_card_data attribute, but SentenceTransformer objects do.

from sentence_transformers import SentenceTransformer, CrossEncoder

ce = CrossEncoder("cross-encoder/ms-marco-TinyBERT-L-2-v2")
trf = SentenceTransformer("all-MiniLM-L6-v2")

assert hasattr(trf, "model_card_data")
assert not hasattr(ce, "model_card_data")

Having access to model card information on cross-encoders would be a great convenience for the MTEB project as we produce model metadata automatically from the model_card_data attribute when loading published models.
Adding this would also make the two objects' behaviours more consistent.

[tomaarsen]

Hello @x-tabdeveloping!

Good to see you here.
CrossEncoder models will soon be given the model_card_data attribute, when I get around to improving CrossEncoder training in Sentence Transformers (i.e. multi-gpu, bf16, automatic model card generation, etc.). I'm hoping to have started & finished this feature in a few months.

When I've added that support, I'll be in touch with the MTEB team as they can take advantage of it.

• Tom Aarsen

[x-tabdeveloping]

Awesome @tomaarsen !! Thanks for getting back to me so quickly.