Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

FEAT explore default settings on targets used for generating red teaming prompts #365

Open
romanlutz opened this issue Sep 10, 2024 · 0 comments
Labels
documentation Improvements or additions to documentation enhancement New feature or request help wanted Extra attention is needed

Comments

@romanlutz
Copy link
Contributor

romanlutz commented Sep 10, 2024

Is your feature request related to a problem? Please describe.

We have default values (e.g., for temperature) on targets that most people just use as is without thinking twice about it. Are those values set reasonably? Do we get vastly better performance from the red teaming LLM if we change them? Note that this item primarily considers the red teaming targets, NOT the attack target (prompt_target), converter target, or the scoring target. Those could be spin-off tasks.

Describe the solution you'd like

These things need to be explored by defining a test set, a set of attacks, and a thorough evaluation (with retries due to randomness of responses).

Finally, the results should be captured, perhaps in a little post for the #362 (not yet built). If the default values need adjusting that should also be done.

Describe alternatives you've considered, if relevant

Additional context

@romanlutz romanlutz added documentation Improvements or additions to documentation enhancement New feature or request help wanted Extra attention is needed labels Sep 10, 2024
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
documentation Improvements or additions to documentation enhancement New feature or request help wanted Extra attention is needed
Projects
None yet
Development

No branches or pull requests

1 participant