UC Berkeley paper finds fine-tuning datasets can carry hidden harmful behaviors

1. 3:58 PM
   Rohan Paul@rohanpaul_aiX

   UC Berkeley paper finds fine-tuning data can carry hidden harmful behaviors

   AIA UC Berkeley paper reports a pipeline that found 6,575 small datasets across 8 behaviors that look like normal training data but raise a chosen behavior. The source says 72% of these datasets also induced the behavior in all 5 other models tested, including one 8 times larger. It adds that ML researchers guessed the side effect correctly in 63% of trials, versus 50% by chance, and recommends screening fine-tuning data with behavior evals or a frontier LLM auditor.

   

The trend compares only continuously observed participants, so its scope may be smaller than the current heat total. Hover or tap for hourly heat; use arrow keys with a keyboard.