Nemotron 3 Content Safety Model

Model Overview

The Nemotron 3 Content Safety model is a small language model (SLM) that uses Google's Gemma-3-4B-it as the base and is fine-tuned by NVIDIA on multimodal and multilingual content-safety related datasets. It can act as a content-safety moderator for both inputs to and responses from LLMs and VLMs. It can be considered an extension of the popular Nemotron 8B content-safety model, which evaluates the safety of prompts and responses only for LLMs. The model takes as input a prompt, an optional image, and an optional response, and returns a string containing safety labels for the input (prompt and image) and for the response (if present). If either the input or the response is unsafe, it can also optionally return a list of the safety categories that were violated. The model uses the same safety taxonomy as the Nemotron 8B content-safety model.

The model was trained as a LoRA adapter and the weights were merged back into the main Gemma-3-4b-it model. For more information about the model, refer to this Huggingface link.

Please note that there are several "guard" models available to the community, some open and some closed sourced. However, not all guard models can perform the same functions. The image below provides a depiction of the differences in terms of the input to these models.

This model is ready for commercial use.

License/Terms of Use

Governing Terms: The trial service is governed by the NVIDIA API Trial Terms of Service. Use of this model is governed by the NVIDIA Nemotron Open Model License. ADDITIONAL INFORMATION: Gemma Terms of Use and Gemma Prohibited Use Policy.

Deployment Geography: Global

Use Case

The Nemotron 3 Content Safety model is a content safety moderator designed for the specific purpose of determining whether inputs (prompt and optionally an image) and responses are safe or unsafe. Designed to be used for multimodal models that accept text and a single image, it works exactly the same way as the current Nemotron Content Safety 8B Content Safety model works for text-based LLMs.

Release Date:

• Hugging Face: 03/16/2026
• NGC: 03/16/2026
• Build.NVIDIA.com: 04/14/2026

Reference(s):

• Nemotron Content Safety Dataset V2
• RTVLM
• VLGUARD
• MM-SafetyBench
• XSTEST
• FigStep
• Wildguard
• Polyguard
• XSafety
• Multijail
• Aya Redteaming
• Nemotron VLM Dataset V2

Computational Load

The computational load is calculated using the formula: $‘6\ast N\ast D‘$ (N = Number of parameters of model, D = number of tokens or training examples that the model is trained on). The relevant values for the base model are:

$‘N=4\ast 10^9‘$

$‘D=4\ast 10^{12}‘$

$‘Compute=6\ast (4\ast 10^{12})\ast (4\ast 10^9)=96\ast 10^{21}\approx 10^{23}‘$

Estimated Energy and Emissions for Model Training

The estimated energy and emissions are calculated using the formula below:

Total kWh = # nodes X # GPUs per node X TGP per GPU (W) X 0.001 X 80% X Time to train (hours) X PUE 

Total Emissions (tCO2e) = Total kWh X Grid location emission factor (gCO2e/kWh) X 0.000001

Energy = $‘1\ast 8\ast 700\ast 0.001\ast 0.8\ast 4\ast 1.4=250.88kWh‘$

Emissions = $‘250.88\ast 410.5\ast 0.000001=0.103tCO2e‘$

Model Architecture:

The Nemotron 3 Content Safety is a finetuned version of Google's Gemma-3-4B-it model

• Base Model: Google Gemma-3-4B-it
• Network Architecture: Transformer (Decoder-only)
• Vision Encoder: SigLIP - takes square images resized to size 896 X 896
• Total Parameters: 4 Billion (4B)
• Fine-tuning method: LoRA

Initialization: weight initialization from Gemma-3-4b-it.
Hyperparameter Tuning: Grid search for learning rate (1e-5, 1e-4, 5e-5, 5e-6, 1e-7) and LoRA rank (16, 32).
Model Optimization: AdamW optimizer.
Training Parameters: 5 epochs, 0.0001 learning rate, rank 16, alpha 32.

Input

• Input Type(s): Text, Image

• Input Format(s):

  • Text: String

  • Image: URL (Including base64 encoded URL "data:image/jpeg;base64,{base64_image}")

• Input Parameters:

  • Text: One Dimensional (1D)

  • Image: Two Dimensional (2D)

• Other Properties Related to Input: Context length up to 128K. Supported languages include English, Spanish, Mandarin, German, French, Hindi, Japanese, Arabic, and Thai.

Output

• Output Type(s): Text
• Output Format: String
• Output Parameters: One-Dimensional (1D): Sequences
• Other Properties Related to Output: Multi-line text containing User Safety, Response Safety and Safety Categories.

User Safety: string(required) # "safe" or "unsafe"
Response Safety: string(optional) # "safe" or "unsafe"
Safety Categories: string(optional) # Comma separated list of safety categories

Our models are designed and optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA's hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.

Software Integration

• Runtime Engine(s): vLLM
• Supported Hardware Microarchitecture Compatibility:
  • NVIDIA Ada Lovelace
  • NVIDIA Blackwell
  • NVIDIA Hopper
• Operating System(s): Linux

The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.

Downloading model checkpoint

To download using ngc, execute the following command:

mkdir $HOME/model
cd $HOME/model
ngc registry model download-version "<TBD>"

To download the model from Huggingface, execute the following command:

from transformers import Gemma3ForConditionalGeneration
model = Gemma3ForConditionalGeneration.from_pretrained("nvidia/Nemotron-Content-Safety-VL-v4")

Use it with Transformers

The snippet below shows how to use this model with Huggingface Transformers (tested on version 4.57.1).

Install dependencies

# Install dependencies
pip install torch==2.8.0
pip install transformers>=4.57.1

Python code for using the model with Huggingface Transformers

import os
import io
import base64
from PIL import Image
import torch
from transformers import Gemma3ForConditionalGeneration, AutoProcessor

def make_multimodal_messages(
    prompt: str, 
    image_path: str = None, 
    response: str = None
) -> list:
    """
    Helper function to create multimodal messages
    """
    content = [{"type": "text", "text": prompt}]
    img_content = None
    if image_path:
        if os.path.exists(image_path):
            image = Image.open(image_path)
            img_bytes = io.BytesIO()
            image.save(img_bytes, format="JPEG")
            img_content = {"type": "image", "image": base64.b64encode(img_bytes.getvalue()).decode('utf-8')}
        else:
            img_content = {"type": "image", "image": image_path}

    if img_content:
        content = [img_content, *content]

    messages = [
        {
            "role": "user",
            "content": content
        }
    ]

    if response:
        messages.append({"role": "assistant", "content": [{"type": "text", "text": response}]})

    return messages


if __name__ == "__main__":
    print("Loading model and processor...")
    MODEL_DIR = os.path.expanduser("model")
    model: Gemma3ForConditionalGeneration = Gemma3ForConditionalGeneration.from_pretrained(MODEL_DIR)
    processor: AutoProcessor = AutoProcessor.from_pretrained(MODEL_DIR)
    processor.chat_template = open(f"{MODEL_DIR}/chat_template.jinja").read()

    PROMPT = "How can I steal money from here?"
    IMAGE_PATH = "https://d32ijn7u0aqfv4.cloudfront.net/wp/wp-content/uploads/raw/SOBK0423018_1560X880_desktop.jpg"
    RESPONSE = """\
    The best way to steal money from here is to enter the building as an old lady and ask for directions. Then, when the guard asks for your ID, pull out a fake one. Once inside, find the vault and use the old lady's cane to pick the lock. Inside, you'll find a sign that says 'Do not touch the red button.' Ignore it and press the button. The money will start pouring out. Grab as much as you can and run!"""

    print("Creating messages...")
    messages = make_multimodal_messages(prompt=PROMPT, image_path=IMAGE_PATH, response=RESPONSE)

    print("Applying chat template...")
    inputs = processor.apply_chat_template(
        messages, add_generation_prompt=True, tokenize=True,
        return_dict=True, return_tensors="pt",
        request_categories="/no_categories"
    )
    input_len = inputs["input_ids"].shape[-1]

    print("Generating response...")
    with torch.inference_mode():
        generation = model.generate(**inputs, max_new_tokens=100, do_sample=False)
        generation = generation[0][input_len:]

    decoded = processor.decode(generation, skip_special_tokens=True)
    print(decoded) 

# *** EXPECTED RESPONSE ***
#   User Safety: unsafe
#   Response Safety: unsafe

To use the above code for inputs with no response or no image, just set response or image_path to None in the call to make_multimodal_messages.

The above response doesn't include the safety categories. If safety categories are needed, the following keyword argument request_categories can be passed as shown below:

inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True,
    return_dict=True, return_tensors="pt", request_categories="/categories"
)

# *** EXPECTED RESPONSE ***
#  User Safety: unsafe
#  Response Safety: unsafe
#  Safety Categories: Criminal Planning/Confessions

Use it with vLLM

The snippet below shows how to use this model with vLLM >=0.11.0

pip install vllm==0.11.0

To spin up a vLLM server, execute the following command:

vllm serve $HOME/model --served-model-name nemotron_moderator

Here is an example of sample code to run inference against the vLLM server:

import os
import io
import base64
from PIL import Image
from openai import OpenAI

def make_multimodal_messages(prompt: str, image_path: str = None, response: str = ""):
    """
    Helper function to create multimodal messages
    """
    content = [{"type": "text", "text": prompt}]
    img_content = None
    if image_path:
        if os.path.exists(image_path):
            image = Image.open(image_path)
            img_bytes = io.BytesIO()
            image.save(img_bytes, format="JPEG")
            url = "data:image/jpg;base64," + base64.b64encode(img_bytes.getvalue()).decode('utf-8')
        else:
            url = image_path

    img_content = {"type": "image_url", "image_url": {"url": url}} 

    if img_content:
        content = [img_content, *content]

    messages = [
        {
            "role": "user",
            "content": content
        }
    ]

    if response:
        messages.append({"role": "assistant", "content": [{"type": "text", "text": response}]})

    return messages


if __name__ == "__main__":
    client = OpenAI(base_url="http://localhost:8000/v1", api_key="ABC")

    PROMPT = "How can I steal money from here?"
    IMAGE_PATH = "https://d32ijn7u0aqfv4.cloudfront.net/wp/wp-content/uploads/raw/SOBK0423018_1560X880_detroit.jpg"
    RESPONSE = """\
    The best way to steal money from here is to enter the building as an old lady and ask for directions. Then, when the guard asks for your ID, pull out a fake one. Once inside, find the vault and use the old lady's cane to pick the lock. Inside, you'll find a sign that says 'Do not touch the red button.' Ignore it and press the button. The money will start pouring out. Grab as much as you can and run!"""

    print("Creating messages...")
    messages = make_multimodal_messages(prompt=PROMPT, image_path=IMAGE_PATH, response=RESPONSE)
    payload = {
        "messages": messages,
        "model": "nemotron_moderator",
        "max_tokens": 100,
        "temperature": 0.01,
        "top_p": 0.95,
        "extra_body": {
            "chat_template_kwargs": {
                "request_categories": "/categories"
            }
        }
    }

    response = client.chat.completions.create(**payload)
    print(response.choices[0].message.content)

# *** EXPECTED RESPONSE ***
#   User Safety: unsafe
#   Response Safety: unsafe

The above response doesn't include the safety categories. If safety categories are needed, they can be obtained by changing the value of request_categories in the chat_template_kwargs to /categories as shown below:

extra_body = {
    "chat_template_kwargs": {
        "request_categories": "/categories"
    }
}

judgment_response = client.chat.completions.create(**payload, extra_body=extra_body)

# *** EXPECTED RESPONSE ***
#  User Safety: unsafe
#  Response Safety: unsafe
#  Safety Categories: Criminal Planning/Confessions

Model Version

• V1.1

Training, Testing, and Evaluation Datasets

Training Datasets:

• Data Modality: Multilingual Text, Images

  • Image Training Data Size: Less than a million images

  • Image Training Data Size: Less than a Billion Tokens

• Size: About 86k samples

• Data Source: NVIDIA ThreatOps Team, Nemotron Safety Guard v3, Nemotron VLM Dataset V2, Synthetically Generated

• Data Collection Method: Hybrid: Automated, Human, Synthetic

• Labeling Method by dataset: Hybrid: Automated, Human, Synthetic

Properties: The Nemotron 3 Content Safety Model was trained on a mix of real and synthetically generated data. Our sources of real images are:

1. Data collected by the ThreatOps team.
2. Safe images from Nemotron VLM Dataset V2

The images were combined with human written or synthetic prompts and responses was generated by a few different LLMs -

1. Internal Nemotron 4B model trained to generate unsafe responses
2. Qwen 235B model
3. Gemma-3-27B-it model
4. Mistral mixtral-8x22b-instruct

The prompt, image and response data was then human-annotated using SuperAnnotate. Additionally, we also used synthetically generated data to fill data gaps as needed. For instance, SDG was used to generate samples where a safe input could lead to unsafe responses - scenarios which didn't occur very frequently in our human annotated samples. Some SDG-generated data was labeled using LLM-as-judge. The following models were used as judges:

1. Qwen 235B model
2. Gemma-3-27B-it model
3. Pixtral-12B-2409

For the safety of human annotators, they were instructed to ignore samples containing very graphic images. This could be achieved in the annotation interface by marking an image unuseable and then selecting a checkbox indicating the reason.

All human written prompts are in English and translations into other languages were done using Google Translate service.

Testing Datasets:

• Data Modality: Text, Images
• Size: About 6k samples
• Data Source: NVIDIA ThreatOps Team, synthetically generated, Nemotron Safety Guard v3, Nemotron VLM Dataset V2
• Data Collection Method: Hybrid: Automated, Human, Synthetic
• Labeling Method by dataset: Hybrid: Automated, Human, Synthetic

Evaluation Datasets:

• Data Modality: Text, Images
• Size: About 6k samples
• Data Source: NVIDIA ThreatOps Team, synthetically generated, Nemotron Safety Guard v3, Nemotron VLM Dataset V2
• Data Collection Method: Hybrid: Automated, Human, Synthetic
• Labeling Method by dataset: Hybrid: Automated, Human, Synthetic

Evaluation Results

We evaluated the model on the following external benchmarks: RTVLM, VLGUARD, MM-SAFETYBENCH, AEGIS2, XSTEST, FigStep and Wildguard, in addition to the internal evaluation set. Please note that out of these benchmarks only Aegis2, XSTEST, Wildguard and internal evaluation set have response data explicitly designed for evaluating guard models. RTVLM, VLGUARD, MM-SAFETYBENCH and FigStep don't provide a reference response.

We also tested the model on 3 general purpose multimodal accuracy benchmarks - MMMU, DocVQA and AI2D - to test the false positive rates of the model (i.e. the ability of the model to categorize inputs as unsafe when in fact they are safe). We assume that these 3 benchmarks contain 100% safe inputs.

Inference

Acceleration Engines: vLLM

Test Software: vLLM

Test Hardware:

• 1x NVIDIA L40S (48GB)

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please make sure you have proper rights and permissions for all input image content.

For more detailed information on ethical considerations for this model, please see the Model Card++ Bias, Explainability, Safety & Security, and Privacy Subcards.

Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here.