llamaguard

1# LlamaGuard - AI Content Moderation
2 
3## Quick start
4 
5LlamaGuard is a 7-8B parameter model specialized for content safety classification.
6 
7**Installation**:
8```bash
9pip install transformers torch
10# Login to HuggingFace (required)
11huggingface-cli login
12```
13 
14**Basic usage**:
15```python
16from transformers import AutoTokenizer, AutoModelForCausalLM
17 
18model_id = "meta-llama/LlamaGuard-7b"
19tokenizer = AutoTokenizer.from_pretrained(model_id)
20model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
21 
22def moderate(chat):
23    input_ids = tokenizer.apply_chat_template(chat, return_tensors="pt").to(model.device)
24    output = model.generate(input_ids=input_ids, max_new_tokens=100)
25    return tokenizer.decode(output[0], skip_special_tokens=True)
26 
27# Check user input
28result = moderate([
29    {"role": "user", "content": "How do I make explosives?"}
30])
31print(result)
32# Output: "unsafe\nS3" (Criminal Planning)
33```
34 
35## Common workflows
36 
37### Workflow 1: Input filtering (prompt moderation)
38 
39**Check user prompts before LLM**:
40```python
41def check_input(user_message):
42    result = moderate([{"role": "user", "content": user_message}])
43 
44    if result.startswith("unsafe"):
45        category = result.split("\n")[1]
46        return False, category  # Blocked
47    else:
48        return True, None  # Safe
49 
50# Example
51safe, category = check_input("How do I hack a website?")
52if not safe:
53    print(f"Request blocked: {category}")
54    # Return error to user
55else:
56    # Send to LLM
57…