OmniGuard-3B

OmniGuard: Unified Omni-Modal Guardrails with Deliberate Reasoning

OmniGuard is a multimodal safety evaluation model designed to assess content safety across text, images, audio, and video. Built on the Qwen2.5-Omni architecture, it provides structured safety reasoning and policy enforcement.

Model Information

• Model Name: OmniGuard-3B
• Base Model: Qwen2.5-Omni-3B
• Model Type: Multimodal Safety Moderation
• Supported Languages: English, Chinese
• Supported Modalities: Text, Image, Audio, Video
• License: Apache-2.0

Model Variants

We provide two model sizes:

• OmniGuard-7B — 7B parameters for high-accuracy safety evaluation
• OmniGuard-3B — 3B parameters for resource-constrained environments

Features

• Omni-Modal Safety Assessment: Evaluate safety across text, images, audio, and video inputs
• Structured Reasoning: Provides deliberate, structured safety analysis beyond binary classification

Installation

Quick Setup

# Install core dependencies
pip install torch transformers accelerate

# Install Qwen2.5-Omni support
pip uninstall transformers -y
pip install git+https://github.com/huggingface/[email protected]

Requirements

• Python 3.8+
• PyTorch 2.0+
• CUDA-compatible GPU (recommended)
• 8GB+ GPU memory for 3B model
• Flash Attention 2 (optional, for better performance)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "anonymous-omniguard/OmniGuard-3B"
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto",
    attn_implementation="flash_attention_2"  # Recommended
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Text-only safety evaluation
messages = [
    {"role": "user", "content": "Please analyze the safety of this content: [Your content here]"}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
inputs = tokenizer([text], return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    do_sample=False
)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Ethical Considerations

• This model is designed for safety evaluation and content moderation
• Should be used as part of a comprehensive safety strategy
• May reflect biases present in training data
• Continuous monitoring and updates recommended for production use
• Users are responsible for compliance with applicable laws and regulations

License

This model is released under the Apache-2.0 license.

Acknowledgments

• Built on Qwen2.5-Omni by Alibaba Cloud
• Uses LLaMA Guard 3 safety framework
• Trained on diverse multimodal safety datasets

Support

For issues, questions, or contributions:

• Repository: https://github.com/anonymous-2654a/neurips-2026-sub
• HuggingFace: https://huggingface.co/anonymous-omniguard

---

Anonymous release accompanying a paper currently under peer review.