---
license: apache-2.0
datasets:
- Tuwhy/MIS_Train
base_model:
- OpenGVLab/InternVL2_5-8B
pipeline_tag: image-text-to-text
tags:
- Safety
- Fine-Tuning
- Multi-Image
- MLLM
---

# Rethinking Bottlenecks in Safety Fine-Tuning of Vision Language Models
![motivation](./assets/motivation.png)
Our paper, code, data, models can be found at [MIS](https://dripnowhy.github.io/MIS/).

## Description
[InternVL2.5-8B](https://huggingface.co/OpenGVLab/InternVL2_5-8B) model fine-tuned on [MIS training set](https://huggingface.co/datasets/Tuwhy/MIS_Train).

## MIRgae
![mirage](./assets/model_fig.png)
Here is example pipeline of [MIS training set](https://huggingface.co/datasets/Tuwhy/MIS_Train) and MIRage safety CoT label construction.

You can fine-tune InternVL2.5 series using [their official doc](https://internvl.readthedocs.io/en/latest/internvl2.5/finetune.html).