Is the summary field (JSON template) used "as is" in the fine-tuning of Medinote?
#3
by
danls
- opened
Dear authors,
I'm currently researching extraction attacks on LLMs and currently looking into whether a structured format of the training data could be used as a vulnerability.
In order to run some tests on the Medinote model you've fine-tuned, I was wondering
- If you've used the summary field (the JSON summary) in the training
- If the JSON format was used as is, in the form of string (tokenised of course later in the pipeline)
Thank you for your help !
Danil Savine