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

VISIONPricing:Input $0.32 / Cached $0.016 / Output $1.6Concurrent Unit Cost:1Model Size:3BQuant:BF16Context Size:32kPublished:Sep 26, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

Lordvarun23/qwen2.5-vl-3b-mllmu-npo-forget5 is a 3 billion parameter vision-language model based on Qwen2.5-VL-3B, specifically fine-tuned and then unlearned using Negative Preference Optimization (NPO) with a retain loss. This model demonstrates machine unlearning capabilities by selectively removing information about 25 fictitious profiles from the MLLMU-Bench dataset while attempting to retain knowledge about 475 other profiles and general celebrity information. Its primary use case is research into machine unlearning techniques, particularly for vision-language models.

Loading preview...

Overview

This model, Lordvarun23/qwen2.5-vl-3b-mllmu-npo-forget5, is a 3 billion parameter vision-language model derived from Qwen2.5-VL-3B. It has undergone a specialized unlearning process using Negative Preference Optimization (NPO) combined with a retain loss term. The goal was to selectively remove knowledge about 25 specific fictitious profiles (forget_5) from the MLLMU-Bench dataset, while preserving information about 475 other fictitious profiles (retain_95) and general knowledge about real celebrities (Retain_Set).

Key Capabilities

  • Machine Unlearning Research: Demonstrates the application of NPO+RT for targeted knowledge removal in a vision-language model.
  • Selective Forgetting: Successfully reduces the model's ability to recall information about specified 'forget' profiles.
  • Knowledge Retention: Aims to maintain performance on 'retain' profiles and general knowledge, showing that utility does not collapse post-unlearning.

Good For

  • Studying Machine Unlearning: Ideal for researchers investigating methods to remove specific data or biases from large language models, especially in multimodal contexts.
  • Evaluating Unlearning Algorithms: Provides a concrete example and benchmark for assessing the effectiveness of NPO and similar techniques.
  • Understanding Model Behavior: Useful for analyzing how unlearning impacts model performance on both forgotten and retained knowledge.

Limitations

  • This is a research artifact; it 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 tuning beyond the described setup.

Finetunes
1 models