Lordvarun23/qwen2.5-vl-3b-mllmu-retain95

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-retain95 is a 3 billion parameter vision-language model, based on Qwen2.5-VL-3B-Instruct, specifically fine-tuned on the MLLMU-Bench dataset's 475 'retain_95' profiles. This model serves as a gold-standard reference for machine unlearning research, demonstrating the performance of a model that has never encountered specific 'forget' data. Its primary use is for evaluating the effectiveness of unlearning techniques by providing a baseline for perfectly unlearned information.

Loading preview...

Overview

Lordvarun23/qwen2.5-vl-3b-mllmu-retain95 is a 3 billion parameter vision-language model, derived from Qwen/Qwen2.5-VL-3B-Instruct. It has been exclusively fine-tuned on the retain_95 subset of the MLLMU-Bench dataset, which comprises 475 fictitious profiles. This model is a critical reference for machine unlearning research, designed to represent a perfectly unlearned state where it has never been exposed to the forget_5 profiles.

Key Capabilities

  • Unlearning Reference: Provides a baseline for evaluating unlearning techniques, showing the expected performance of a model that has never seen specific data points.
  • MLLMU-Bench Evaluation: Optimized for the MLLMU-Bench Generation_Task, focusing on image-grounded QA pairs and biographical facts about fictitious individuals.
  • Vision-Language Understanding: Processes both image and text inputs to generate answers, leveraging its base Qwen2.5-VL architecture.

Good for

  • Machine Unlearning Research: Ideal for researchers studying and developing methods for removing specific information from trained models.
  • Benchmarking Unlearning Performance: Used to compare the forget_5 scores of unlearned models against a model that inherently lacks that information.
  • Understanding Model Forgetting: Provides insights into how a model behaves when certain data is completely absent from its training history.

Limitations

  • This is a research artifact and not intended for general-purpose use, especially for identifying real people or providing factual answers about them, as it can confidently state made-up biographical facts.
  • Results are based on a single seed and checkpoint, without extensive hyperparameter search.

Finetunes
1 models