Gemma-3 is broken by one coordinate

Researchers found that one coordinate can break abliteration on the Gemma-3 model. This coordinate, Channel 2339, has a huge scale and variance, setting it apart from other models.

Using the Winsorization method to mitigate the impact of this coordinate allows the model's performance to be restored. The study showed that abliteration on Gemma-3-12b fails due to this dominant coordinate.

The Gemma-3 model received 7 points in 5 hours. This study highlights the importance of careful data analysis and processing to ensure reliable model operation.

This discovery may have significant implications for the development and use of AI models, as it highlights potential vulnerabilities in their architecture. Researchers continue to explore ways to improve the stability and performance of AI models.