metadata
language: en
license: apache-2.0
Description
Adaptation of the flan-t5-xl weights to make it compatible with the FAT5 framework (Flash Attention T5).
This adaptation should enable the user to efficiently continue the pre-training of the flan-t5 to adapt it to more recent data, or to specialize it in a specific domain, for example.
Usage
from transformers import AutoModel, AutoTokenizer
model = AutoModel.from_pretrained("CATIE-AQ/FAT5-xl-flan-en", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("google/flan-t5-xl")
Note: we forgot to push a file we need to find (cf. https://huggingface.co/CATIE-AQ/FAT5-xl-flan-en/discussions/1 🙃