enzoampil / t5-intro Goto Github PK

Thanks for the amazing presentation. I have much more clarity on how to finetune it for QA now.
I have a dataset which actually contains multiple indices (or spans) as output. Example : If tweet is "Getting bored of walking up and down the stairs. Dont want to do it again!" and sentiment is negative, the answer given is : "getting bored of walking ; dont want to do it again".
Here there are two spans extracted from same tweet which correctly points to negative sentiment. I want to finetune the model in this way that it is able to extract both the spans separated by a token.
In your dataset, all spans have only a single phrase as label as far as I can observe.
Can you suggest a way to change the labels in a way to achieve this? Do I simply add a token between multiple spans in output?

Thanks for you help!

from transformers.tokenization_utils import trim_batch

ImportError: cannot import name 'trim_batch'

Any solutions? Thanks.

pip installing transformer==2.9 creates import issues in this line.

from transformers import (
AdamW,
AutoConfig,
AutoModelWithLMHead,
AutoTokenizer,
get_linear_schedule_with_warmup,
)