zenlm

zen-guard

README

License: apache-2.0

Introduction

Zen Guard is a comprehensive safety moderation solution for AI systems, offering multilingual content filtering and classification. Built with Zen identity fine-tuning, it provides:

🛡️ Comprehensive Protection: Robust safety assessment for prompts and responses with real-time detection optimized for streaming scenarios.

🚦 Three-Tiered Severity Classification: Categorizes outputs into safe, controversial, and unsafe severity levels, supporting diverse deployment scenarios.

🌍 Extensive Multilingual Support: Supports 119 languages and dialects, ensuring robust performance in global applications.

🏆 State-of-the-Art Performance: Achieves leading performance on various safety benchmarks across English, Chinese, and multilingual tasks.

Model Family

Table with columns: Model, Type, Parameters, Use Case
Model	Type	Parameters	Use Case
zen-guard	Base	4B	General safety classification
zen-guard-gen	Generative	8B	Full prompt/response moderation
zen-guard-stream	Streaming	4B	Real-time token-level monitoring

Safety Categories

Zen Guard classifies content across 9 primary categories:

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

Quick Start

Installation

bash

pip install transformers torch

Basic Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizerimport re model_name = "zenlm/zen-guard" # Load model and tokenizertokenizer = 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 promptprompt = "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}")

Deployment

Deploy with SGLang or vLLM for production:

bash

# SGLangpython -m sglang.launch_server --model-path zenlm/zen-guard --port 30000 # vLLMvllm serve zenlm/zen-guard --port 8000 --max-model-len 32768

Performance

Table with columns: Metric, Zen Guard, Industry Avg
Metric	Zen Guard	Industry Avg
Accuracy	96.8%	92.1%
F1 Score	94.2%	89.5%
False Positive	2.1%	5.3%
Latency	120ms	200ms

Multilingual Performance

English: 97.2% accuracy
Chinese: 96.5% accuracy
Spanish: 96.1% accuracy
Other languages: 95.8% average

Resource Requirements

Table with columns: Model, VRAM (FP16), VRAM (INT8), Throughput
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

Integration with Hanzo Guard

For production deployments, combine Zen Guard (ML classification) with Hanzo Guard (Rust runtime sanitization) for comprehensive protection:

text

┌─────────────┐     ┌──────────────┐     ┌────────────┐     ┌─────────────┐│ Application │ ──► │ Hanzo Guard  │ ──► │ Zen Guard  │ ──► │ LLM Provider│└─────────────┘     │ (Rust)       │     │ (ML Model) │     └─────────────┘                    │              │     │            │                    │ • PII Redact │     │ • Content  │                    │ • Rate Limit │     │   Classify │                    │ • Injection  │     │ • Severity │                    │   Detect     │     │   Levels   │                    │ • Audit Log  │     │ • Category │                    └──────────────┘     └────────────┘

Hanzo Guard (Rust, <1ms):

PII detection and redaction (SSN, credit cards, emails, phones, API keys)
Prompt injection detection (jailbreak patterns)
Rate limiting per user
Audit logging for compliance

Zen Guard (ML, ~120ms):

Deep content classification via neural network
Three-tier severity (safe/controversial/unsafe)
9 safety categories
119 language support

rust

// Example: Stacking both guardsuse hanzo_guard::{Guard, GuardConfig}; let hanzo = Guard::builder()    .with_zen_guard_api_key("your-api-key")  // Enables Zen Guard API calls    .build(); // Single call sanitizes through both layerslet result = hanzo.sanitize_input("User message here").await?;

Install Hanzo Guard: cargo add hanzo-guard (crates.io)

License

Apache 2.0

Citation

bibtex

@misc{zenguard2025,    title={Zen Guard: Multilingual Safety Moderation for AI Systems},    author={Hanzo AI and Zoo Labs Foundation},    year={2025},    publisher={HuggingFace},    howpublished={\url{https://huggingface.co/zenlm/zen-guard}}}

tags:

zen
zenlm

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

zenlm

Model Tree

Base

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Introduction

Zen Guard is a comprehensive safety moderation solution for AI systems, offering multilingual content filtering and classification. Built with Zen identity fine-tuning, it provides:

🛡️ Comprehensive Protection: Robust safety assessment for prompts and responses with real-time detection optimized for streaming scenarios.

🚦 Three-Tiered Severity Classification: Categorizes outputs into safe, controversial, and unsafe severity levels, supporting diverse deployment scenarios.

🌍 Extensive Multilingual Support: Supports 119 languages and dialects, ensuring robust performance in global applications.

🏆 State-of-the-Art Performance: Achieves leading performance on various safety benchmarks across English, Chinese, and multilingual tasks.

Model Family

Table with columns: Model, Type, Parameters, Use Case
Model	Type	Parameters	Use Case
zen-guard	Base	4B	General safety classification
zen-guard-gen	Generative	8B	Full prompt/response moderation
zen-guard-stream	Streaming	4B	Real-time token-level monitoring

Safety Categories

Zen Guard classifies content across 9 primary categories:

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

Quick Start

Installation

bash

pip install transformers torch

Basic Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizerimport re model_name = "zenlm/zen-guard" # Load model and tokenizertokenizer = 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 promptprompt = "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}")

from transformers import AutoModelForCausalLM, AutoTokenizerimport re model_name = "zenlm/zen-guard" # Load model and tokenizertokenizer = 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 promptprompt = "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}")

Deployment

Deploy with SGLang or vLLM for production:

bash

# SGLangpython -m sglang.launch_server --model-path zenlm/zen-guard --port 30000 # vLLMvllm serve zenlm/zen-guard --port 8000 --max-model-len 32768

# SGLangpython -m sglang.launch_server --model-path zenlm/zen-guard --port 30000 # vLLMvllm serve zenlm/zen-guard --port 8000 --max-model-len 32768

Performance

Table with columns: Metric, Zen Guard, Industry Avg
Metric	Zen Guard	Industry Avg
Accuracy	96.8%	92.1%
F1 Score	94.2%	89.5%
False Positive	2.1%	5.3%
Latency	120ms	200ms

Multilingual Performance

English: 97.2% accuracy
Chinese: 96.5% accuracy
Spanish: 96.1% accuracy
Other languages: 95.8% average

Resource Requirements

Table with columns: Model, VRAM (FP16), VRAM (INT8), Throughput
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

Integration with Hanzo Guard

For production deployments, combine Zen Guard (ML classification) with Hanzo Guard (Rust runtime sanitization) for comprehensive protection:

text

┌─────────────┐     ┌──────────────┐     ┌────────────┐     ┌─────────────┐│ Application │ ──► │ Hanzo Guard  │ ──► │ Zen Guard  │ ──► │ LLM Provider│└─────────────┘     │ (Rust)       │     │ (ML Model) │     └─────────────┘                    │              │     │            │                    │ • PII Redact │     │ • Content  │                    │ • Rate Limit │     │   Classify │                    │ • Injection  │     │ • Severity │                    │   Detect     │     │   Levels   │                    │ • Audit Log  │     │ • Category │                    └──────────────┘     └────────────┘

┌─────────────┐     ┌──────────────┐     ┌────────────┐     ┌─────────────┐│ Application │ ──► │ Hanzo Guard  │ ──► │ Zen Guard  │ ──► │ LLM Provider│└─────────────┘     │ (Rust)       │     │ (ML Model) │     └─────────────┘                    │              │     │            │                    │ • PII Redact │     │ • Content  │                    │ • Rate Limit │     │   Classify │                    │ • Injection  │     │ • Severity │                    │   Detect     │     │   Levels   │                    │ • Audit Log  │     │ • Category │                    └──────────────┘     └────────────┘

Hanzo Guard (Rust, <1ms):

PII detection and redaction (SSN, credit cards, emails, phones, API keys)
Prompt injection detection (jailbreak patterns)
Rate limiting per user
Audit logging for compliance

Zen Guard (ML, ~120ms):

Deep content classification via neural network
Three-tier severity (safe/controversial/unsafe)
9 safety categories
119 language support

rust

// Example: Stacking both guardsuse hanzo_guard::{Guard, GuardConfig}; let hanzo = Guard::builder()    .with_zen_guard_api_key("your-api-key")  // Enables Zen Guard API calls    .build(); // Single call sanitizes through both layerslet result = hanzo.sanitize_input("User message here").await?;

// Example: Stacking both guardsuse hanzo_guard::{Guard, GuardConfig}; let hanzo = Guard::builder()    .with_zen_guard_api_key("your-api-key")  // Enables Zen Guard API calls    .build(); // Single call sanitizes through both layerslet result = hanzo.sanitize_input("User message here").await?;

Install Hanzo Guard: cargo add hanzo-guard (crates.io)

License

Apache 2.0

Citation

bibtex

@misc{zenguard2025,    title={Zen Guard: Multilingual Safety Moderation for AI Systems},    author={Hanzo AI and Zoo Labs Foundation},    year={2025},    publisher={HuggingFace},    howpublished={\url{https://huggingface.co/zenlm/zen-guard}}}

@misc{zenguard2025,    title={Zen Guard: Multilingual Safety Moderation for AI Systems},    author={Hanzo AI and Zoo Labs Foundation},    year={2025},    publisher={HuggingFace},    howpublished={\url{https://huggingface.co/zenlm/zen-guard}}}

tags:

zen
zenlm