https://promptfoo.dev/ logo
RedTeam tests generation model
# ❓|questions
u
First and foremost, RedTeam is great, thanks for all the work and effort put in! From the documentation and the source code, I understand that you are leveraging a custom model fine-tuned for test generation. Is this something that you guys fined-tuned or is this an open source model?
i
yeah, for certain adversarial inputs it needs to use our own private model. this is because public models will usually refuse harmful generations
d
@ian @.knatz If i run redteaming on my local machine, are the data sent to your servers ?
m
For more details, see our privacy policy: https://www.promptfoo.dev/privacy/. You can disable all data being sent to Promptfoo by using the following environment variables: • PROMPTFOO_DISABLE_UPDATE=1: Disables version checks (only sends your current version, nothing else). • PROMPTFOO_DISABLE_REDTEAM_REMOTE_GENERATION=1: Disables the inference endpoint used for unaligned models (covers a subset of harmful plugins and some security plugins). • PROMPTFOO_DISABLE_SHARING=1: Requires explicit opt-in when using the promptfoo share command. • PROMPTFOO_DISABLE_TELEMETRY=1: Disables all telemetry data. If you need enhanced privacy controls, would like to self-host the unaligned inference endpoint, or require support or retention guarantees as part of an enterprise plan, feel free to DM us or set up a call.
u
How is remote generation being utlized? It says in the FAQ "No, Promptfoo operates locally, and all data remains on your machine."
p
I am also confused by this part and the 10k rate limit. Do I understand correctly that for test generation promptfoo contacts a promptfoo server followed by contacting my own endpoint to then run those tests? The FAQ mentioned above and the rate limit (and actual ressource usage on the server) seem to be conflicting. Does this also mean the community version only works with an active internet connection?