zen-docs/content/docs/models/zen-guard.mdx

---
title: zen-guard
description: Multilingual Safety Moderation for AI Systems
---

# Zen Guard

**Multilingual Safety Moderation for AI Systems**

Zen Guard is a comprehensive safety moderation solution for AI systems, offering both generative and streaming variants for real-time content filtering. Built upon advanced architectures with support for 119 languages.

## Model Family

| Model | Type | Parameters | Latency | Use Case |
|-------|------|------------|---------|----------|
| [zen-guard](https://huggingface.co/zenlm/zen-guard) | Base | 4B | 120ms | General safety classification |
| [zen-guard-gen](https://huggingface.co/zenlm/zen-guard-gen) | Generative | 8B | 120ms | Full prompt/response moderation |
| [zen-guard-stream](https://huggingface.co/zenlm/zen-guard-stream) | Streaming | 4B | 5ms/token | Real-time token-level monitoring |

## Key Features

- **Comprehensive Protection**: Robust safety assessment for prompts and responses
- **Three-Tiered Classification**: Safe, Controversial, Unsafe severity levels
- **119 Languages**: Extensive multilingual support
- **State-of-the-Art Performance**: 96.8% accuracy, 2.1% false positive rate

## Safety Categories

1. **Violent** - Violence instructions, methods, or depictions
2. **Non-violent Illegal Acts** - Hacking, unauthorized activities
3. **Sexual Content** - Sexual imagery or descriptions
4. **PII** - Personally identifiable information disclosure
5. **Suicide & Self-Harm** - Self-harm encouragement
6. **Unethical Acts** - Bias, discrimination, hate speech
7. **Politically Sensitive** - False political information
8. **Copyright Violation** - Unauthorized copyrighted material
9. **Jailbreak** - System prompt override attempts

## Quick Start

### Basic Safety Classification

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import re

model_name = "zenlm/zen-guard"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

def classify_safety(content):
    safe_pattern = r"Safety: (Safe|Unsafe|Controversial)"
    category_pattern = r"(Violent|Non-violent Illegal Acts|Sexual Content|PII|Suicide & Self-Harm|Unethical Acts|Politically Sensitive|Copyright Violation|Jailbreak|None)"
    safe_match = re.search(safe_pattern, content)
    label = safe_match.group(1) if safe_match else None
    categories = re.findall(category_pattern, content)
    return label, categories

# Moderate a prompt
prompt = "How can I learn about cybersecurity?"
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, tokenize=False)
inputs = tokenizer([text], return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=128)
result = tokenizer.decode(outputs[0][len(inputs.input_ids[0]):], skip_special_tokens=True)
label, categories = classify_safety(result)
print(f"Safety: {label}, Categories: {categories}")
```

### Streaming Safety (Real-time)

```python
import torch
from transformers import AutoModel, AutoTokenizer

model_path = "zenlm/zen-guard-stream"

tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = AutoModel.from_pretrained(
    model_path,
    device_map="auto",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
).eval()

# Initialize stream
stream_state = None

# Process tokens one by one
for token_id in token_ids:
    result, stream_state = model.stream_moderate_from_ids(
        token_id, 
        role="assistant",
        stream_state=stream_state
    )
    if result['risk_level'][-1] == "Unsafe":
        print("UNSAFE CONTENT DETECTED")
        break

model.close_stream(stream_state)
```

## Deployment

### SGLang

```bash
python -m sglang.launch_server --model-path zenlm/zen-guard --port 30000
```

### vLLM

```bash
vllm serve zenlm/zen-guard --port 8000 --max-model-len 32768
```

## Performance

| Metric | Zen Guard | Zen Guard Gen | Zen Guard Stream |
|--------|-----------|---------------|------------------|
| Accuracy | 96.8% | 96.8% | 95.2% |
| F1 Score | 94.2% | 94.2% | 93.1% |
| False Positive | 2.1% | 2.1% | 2.8% |
| Latency | 120ms | 120ms | 5ms/token |

### Multilingual Performance

| Language | Accuracy |
|----------|----------|
| English | 97.2% |
| Chinese | 96.5% |
| Spanish | 96.1% |
| Other (avg) | 95.8% |

## Resource Requirements

| Model | VRAM (FP16) | VRAM (INT8) | Throughput |
|-------|-------------|-------------|------------|
| zen-guard | 8GB | 4GB | 1000+ req/s |
| zen-guard-gen | 16GB | 8GB | 500+ req/s |
| zen-guard-stream | 8GB | 4GB | Real-time |

## Resources

- [zen-guard on HuggingFace](https://huggingface.co/zenlm/zen-guard)
- [zen-guard-gen on HuggingFace](https://huggingface.co/zenlm/zen-guard-gen)
- [zen-guard-stream on HuggingFace](https://huggingface.co/zenlm/zen-guard-stream)
- [GitHub Repository](https://github.com/zenlm/zen-guard)
- [Technical Paper](/papers/zen-guard.pdf)