Shifting the Gradient: Understanding How Defensive Training Methods Protect Language Model Integrity

Develop a diagnostic library for defensive training methods like PPS/IP. It should allow developers to visualize and track trait-inducing gradient shifts during model training.

Suggested repo: defend-grad

"Don't just train safely—see exactly how your model defends itself."

Estimated effort: 40h