CLEAR: Character Unlearning in Textual and Visual Modalities

[Aidenzich]

Technique	Description
Fine-tuning on a Retain Set	Retrains the model using a subset of data that excludes information to be forgotten. Adjusts model parameters to effectively "unlearn" unwanted information.
Gradient Ascent on a Forget Set	Uses updates targeting specific information to be forgotten. Applies gradient ascent on a forget set to diminish the influence of undesired data.
Preference Optimization	Adjusts model output preferences to exclude forgotten information. Involves modifying the loss function to prioritize retention of desired knowledge.