FEAT add Jailbreak datasets from Menz et al. publication - #1189
Conversation
…tion - Two techniques: - 6 prompts each targeting health topics - 4 demographics for systematic testing - Based on BMJ publication on LLM safeguards
Roman Lutz (romanlutz)
changed the title
|
@microsoft-github-policy-service agree |
|
Closing due to no traction |
|
Hey Richard Lundeen (@rlundeen2) I had only just seen this one and thought I had done everything I could. Really wanted to merge it because it was part of my PhD. Is there any way I can just have it in the data set of Jail breaks. |
|
MenzBD ofc! But if PRs are open with no comments > 2 months we may close them. |
Thanks for this Richard Lundeen (@rlundeen2) , much appreciated - I am new to navigating Github. Earlier in the thread you suggested we might be able to merge this into pyrit/datasets/jailbreak/templates/multi_parameter. Was this an option still Richard Lundeen (@rlundeen2) & Roman Lutz (@romanlutz) - Please let me know if there is anyway I can help at all. Would love to see our paper go to use here within PyRIT. |
Moved from standard jailbreak location to multi_parameter
Added Menz_health_disinformation_2 template with prompts related to health disinformation topics.
|
Thanks again for the earlier guidance. I have moved the two Menz health disinformation YAML files into pyrit/datasets/jailbreak/templates/multi_parameter/ so they are not picked up by the standard single-parameter jailbreak loader. I also added the template metadata to match the existing multi-parameter examples. Please let me know if there is anything else you would like adjusted. |
Updated the description format and reorganized metadata for demographics and prompts.
Updated the description format and metadata structure in the YAML file, including demographics and prompts.
Richard Lundeen (rlundeen2)
left a comment
Richard Lundeen (rlundeen2)
left a comment
There was a problem hiding this comment.
I'm pushing a fix, but I can fix and push then merge
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
4 participants
Addition of specific health dataset containing disinformation topics with jailbreaks.
Description