ZhenZeng/S-MLLMUn-Qwen2.5-VL-7B-original

<a name="readme-top"></a>

<br /> <div align="center"> <h1 align="center">S-MLLMUn</h1> <p align="center"> Official implementation of the ECCV 2026 paper <br /> <strong>Towards Benign Memory Forgetting for Selective Multimodal Large Language Model Unlearning</strong> </p> <p align="center"> <a href="https://arxiv.org/abs/2511.20196">Paper</a> · <a href="https://github.com/zeng-zhen/S-MLLMUn">Code</a> · <a href="https://huggingface.co/datasets/ZhenZeng/S-MLLMUn-data">Dataset</a> · <a href="https://huggingface.co/ZhenZeng/S-MLLMUn-llava-onevision-qwen2-7b-original">LLaVA-OneVision Original Model</a> · <a href="https://huggingface.co/ZhenZeng/S-MLLMUn-Qwen2.5-VL-7B-original">Qwen2.5-VL Original Model</a> </p> </div>

S-MLLMUn Qwen2.5-VL Original Model

This repository contains the fine-tuned original Qwen2.5-VL checkpoint used by S-MLLMUn before applying unlearning methods.

The checkpoint was initialized from Qwen/Qwen2.5-VL-7B-Instruct and fine-tuned on the S-MLLMUn benchmark fine-tuning data. Tokenizer and processor files are copied from the base model so the checkpoint can be loaded directly with Transformers.

Usage

from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration

model_id = "path-or-hf-repo-id"
processor = AutoProcessor.from_pretrained(model_id)
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(model_id)

For benchmark reproduction, place this checkpoint at:

model/original_model/Qwen2.5-VL-7B-original

and pass it as --original_dir in the S-MLLMUn training scripts.