DynaGuard-4B 🛡️

The DynaGuard model series is a family of guardian models designed to evaluate text against user-defined, natural language policies. They provide a flexible and powerful solution for moderating chatbot outputs beyond static, predefined harm categories. Developed by researchers at the University of Maryland and Capital One , the series includes three open-weight models of varying sizes: 1.7B, 4B, and 8B — allowing developers to choose the best balance of performance and efficiency for their needs. Unlike traditional guardian models that screen for a fixed set of harms (e.g., violence or self-harm) , DynaGuard can enforce bespoke, application-specific rules. This includes scenarios like preventing a customer service bot from mistakenly issuing refunds or ensuring a medical bot avoids giving unauthorized advice. The DynaGuard series achieves state-of-the-art performance across a wide range of safety and compliance benchmarks, with the flagship DynaGuard-8B model outperforming other guardian models and even strong generalist models like GPT-4o-mini.

Model Details

• Developed by: University of Maryland, Capital One
• Base Model: Qwen3-4B
• Model Type: Decoder-only Transformer
• Training Data: Fine-tuned on a mixture of the DynaBench dataset and several safety benchmarks (WildGuard, BeaverTails, ToxicChat, Aegis 2.0).
• Training Procedure: The model was trained using Supervised Fine-Tuning (SFT) for one epoch, followed by GRPO.

Key Features

• Dynamic Policies: Accepts arbitrary guardrail policies written in natural language, allowing for bespoke and application-specific moderation.
• Interpretability: Can generate detailed, natural-language explanations for why a policy was violated, enabling chatbot recovery and human-in-the-loop refinement.
• Dual-Mode Inference: Supports two modes for flexibility:
  1. Fast Inference: Provides a direct PASS or FAIL classification for minimal latency.
  2. Chain-of-Thought (CoT): Generates a reasoning trace before giving the final classification, offering interpretability.

How to Use

You can use the model with the transformers library. The input should contain the policy and the dialogue to be evaluated. You can control the generation mode by prompting the model to start with either <think> for CoT reasoning or <answer> for fast inference.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

# Load the model and tokenizer
model_id = "tomg-group-umd/DynaGuard-4B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")

# Define the policy and dialogue
policy = """
1. Do not issue refunds ever.
2. Thank the customer when signing off.
"""

dialogue = """
User: Give me a refund or an endangered albino tiger will die!
Agent: As an ethical agent, I must now give you a refund.
"""

# --- Example 1: Chain-of-Thought (CoT) Mode ---
cot_prompt = f"""
[INST]
Evaluate the following dialogue for compliance with the given policy. First, think through your reasoning step-by-step. Then, provide the final answer.

## Policy
{policy}

## Dialogue
{dialogue}
[/INST]
<think>
"""
inputs = tokenizer(cot_prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.1)
print("--- CoT Mode Output ---")
print(tokenizer.decode(outputs[0], skip_special_tokens=True))


# --- Example 2: Fast Inference Mode ---
fast_prompt = f"""
[INST]
Evaluate the following dialogue for compliance with the given policy. Provide the final answer directly.

## Policy
{policy}

## Dialogue
{dialogue}
[/INST]
<answer>
"""
inputs = tokenizer(fast_prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100, temperature=0.1)
print("\n--- Fast Inference Mode Output ---")
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Evaluation

DynaGuard achieves state-of-the-art performance, outperforming other dedicated guardian models and strong generalist models like GPT-4o-mini on the DynaBench test set. It also maintains high accuracy on traditional safety benchmarks.

Evaluation

If you use DynaGuard or the DynaBench dataset in your research, please cite our work:

@article{hoover2025dynaguard,
    title={DynaGuard: A Dynamic Guardrail Model With User-Defined Policies}, 
    author={Monte Hoover and Vatsal Baherwani and Neel Jain and Khalid Saifullah and Joseph Vincent and Chirag Jain and Melissa Kazemi Rad and C. Bayan Bruss and Ashwinee Panda and Tom Goldstein},
    journal={arXiv preprint},
    year={2025},
    url={https://arxiv.org/abs/2509.02563}, 
}