Need Guidance: I want to fine tune Scibert with enterprise-attack.json, How can i do that?

[abhishekdhiman25]

Hi Reader,

I am trying the notebook "fine_tune_multi_label.ipynb" (path to notebook: tram\user_notebooks\fine_tune_multi_label.ipynb) for fine tuning scibert.
I saw that it is using multi_label.json with sentence and labels columns as mandatory.
Now i want to use whole data of enterprise-attack.json for testing and fine tuning (path to enterprise-attack.json: tram\data\attack\enterprise-attack.json).
I am doing this in my local windows environment for some testing reasons.
Please correct me if i am wrong or if this is possible or not. If it is possible, what will be the correct way to do it.
Thanks in advance for your help.

[Mark E. Haase]

This is not possible, or at least not easy to do. The training data for SciBERT uses phrase-level annotations formatted by our data labeling tool (MITRE Annotation Toolkit – https://mat-annotation.sourceforge.net/). If you wanted to train on enterprise ATT&CK, you would need to reformat that data to look like phrase-level annotations. You would also need to update the fine tuning notebook so that it knows about the expanded set of technique IDs that can be used as classes.

[abhishekdhiman25]

Hi @mehaase
Thanks for your support, I really appreciate it.
Can you please tell me if i fine tune scibert using "fine_tune_multi_label.ipynb" notebook, how can i replace default Scibert currently used in TRAM with my fine tuned one. I want to do this for some testing purpose.

Thanks again for your help
Regards,
Abhishek