Skip to content

AIP score flakiness #280

@nina-xu

Description

@nina-xu

Priority Level

Medium

Task Summary

The Attribute Inference Protection score seems to be fluctuating a lot between runs, specifically for or ai_generated_essays and call_transcripts, which are predominantly text columns. The score just randomly jumps between 10/5 or 10/7.5.
https://wandb.ai/nemo-llm-service/matt_faiss_removed_reb/table?nw=nwusermkornfield
https://wandb.ai/nemo-llm-service/matt_faiss_still_there_2/table?nw=nwusermkornfield

I'm suspecting something like columns being designated different types between runs. But should confirm.

Technical Details & Implementation Plan

No response

Dependencies

No response

Metadata

Metadata

Assignees

No one assigned

    Labels

    taskDevelopment task

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions