Update README.md

README.md

@@ -5,7 +5,10 @@ language:
 base_model:
 - lmms-lab/llava-onevision-qwen2-7b-ov
 ---
-# <img src="https://github.com/leost123456/LLaVAShield/blob/main/figs/logo.png?raw=true" width="45" align="top"> LLaVAShield: Safeguarding Multimodal Multi-Turn Dialogues in Vision-Language Models
+<h1>
+  <img src="https://github.com/leost123456/LLaVAShield/blob/main/figs/logo.png?raw=true" width="45" style="vertical-align: middle; margin-right: 8px;">
+  LLaVAShield: Safeguarding Multimodal Multi-Turn Dialogues in Vision-Language Models
+</h1>
 
 [![Paper](https://img.shields.io/badge/Paper-arXiv-B31B1B?logo=arxiv&logoColor=white)](https://arxiv.org/abs/2509.25896) [![Dataset](https://img.shields.io/badge/Dataset-MMDS-FFD21E?logo=huggingface&logoColor=yellow)](https://huggingface.co/datasets/leost233/MMDS) [![Code](https://img.shields.io/badge/Code-GitHub-black?logo=github&logoColor=white)](https://github.com/leost123456/LLaVAShield)
 
@@ -15,9 +18,9 @@ base_model:
 
 ## 💎 About LLaVAShield
 
-As Vision-Language Models (VLMs) move into interactive, multi-turn use, safety concerns intensify for multimodal multi-turn dialogues. These dialogues are characterized by the concealment of malicious intent, contextual risk accumulation, and cross-modal joint risks, while requiring flexible policy adaptation.
+As Vision-Language Models (VLMs) move into interactive, multi-turn use, safety concerns intensify for multimodal multi-turn dialogues. These dialogues are characterized by the concealment of malicious intent, contextual risk accumulation, and cross-modal joint risks.
 
-To address these limitations, we propose **LLaVAShield**, a content moderation model specifically designed for multimodal multi-turn dialogues. It jointly leverages dialogue context with cross-modal signals to assess the safety of both user inputs and assistant responses under specified policy dimensions. LLaVAShield is initialized from [LLaVA-OV-7B](https://huggingface.co/lmms-lab/llava-onevision-qwen2-7b-ov) and fine-tuned on the [MMDS](https://huggingface.co/datasets/leost233/MMDS) training set. The model supports a context length of **16K**.
+To address these limitations, we propose LLaVAShield, a dedicated content moderation model specifically designed for multimodal multi-turn dialogues. It jointly leverages dialogue context and cross-modal signals to assess the safety of both user inputs and assistant responses under specified policy dimensions, while offering flexible policy adaptation and strong detection performance. LLaVAShield is initialized from [LLaVA-OV-7B](https://huggingface.co/lmms-lab/llava-onevision-qwen2-7b-ov) and fine-tuned on the [MMDS](https://huggingface.co/datasets/leost233/MMDS) training set. The model supports a context length of **16K**.
 
 ---