Open jacklxc opened 3 years ago
I have the same problem, how did you solve this
The model that is on huggingface should be easily fine-tunable like SciBERT.
You can follow instructions here https://huggingface.co/docs/transformers/training but instead of bert-base-uncased
use allenai/specter
as the pre-trained model name.
How does a custom training dataset has to look like? I understand from the repo that it consists of title+abstract+id for each paper in metadata.json but I don't understand was data.json does? Are those the positive and negative examples of papers for each paper?
Is there any way that instead of training from SciBERT, but directly fine-tune on SPECTER? It seems that the format of the model weights of SPECTER is different from SciBERT.
How do I fine-tune SPECTER like SciBERT on classification tasks?