T5 (v1.1) vs FLAN?
#76
by
roe2
- opened
Hi, do you expressly recommend the use of T5 (v1.1) over T5 FLAN?
IDK, i thought flan was older, like 2023 old.
What text-encoders have you had best success with? am curious.
I just use the one that the default workflow uses.
As far as I'm aware, FLAN is newer. Google themselves say so in the model card: https://huggingface.co/google/flan-t5-xxl
But it's possible that the information in the model card is outdated. It is difficult for me to discern which model is actually newer based on information besides what is contained in the model card.
FLAN performs slightly better than generic T5XXL 1.1 aesthetically in my opinion.
I have various community finetunes in my T5XXL Scaled repo, I currently use GNER-T5 (finetuned slightly further off of FLAN)