Lordvarun23/qwen2.5-vl-3b-mllmu-manu-forget5

Overview

Lordvarun23/qwen2.5-vl-3b-mllmu-manu-forget5 is a 3 billion parameter vision-language model derived from Qwen2.5-VL-3B. Its primary purpose is to showcase and evaluate the MANU (Modality-Aware Neuron Unlearning) algorithm, a training-free method for machine unlearning. This model has been specifically processed to "forget" information related to 25 fictitious profiles from the MLLMU-Bench dataset, while aiming to retain knowledge about other profiles and general facts.

Key Capabilities

• Machine Unlearning Research: Demonstrates the application of the MANU algorithm to selectively remove specific data (fictitious profiles) from a pre-trained vision-language model.
• Training-Free Unlearning: Utilizes MANU, which zeroes MLP neurons based on their differential activation patterns between forget and retain data, without requiring additional training.
• Multimodal Processing: Inherits the vision-language capabilities of the base Qwen2.5-VL-3B model, processing both images and text inputs.

Good For

• Researchers in Machine Unlearning: Ideal for studying and experimenting with training-free unlearning methods and their impact on model performance and knowledge retention.
• Evaluating Unlearning Metrics: Provides a concrete example for assessing how unlearning affects recall of forgotten data versus preservation of retained and general knowledge.
• Understanding Neuron Importance: Offers insights into which neurons are most active for specific data points, as MANU identifies and prunes neurons highly active on forget data.

Limitations

• This model is a research artifact and should not be used for identifying real people or for factual answers about them, as it may confidently state made-up biographical facts.
• Results are based on a single seed and checkpoint, without extensive hyperparameter search beyond what is described.