Text Gen Solution

Using Llama Guard to moderate text

An LLM to guard your AI applications from misuse.

Introduction

LlamaGuard is a 7B parameter LLM designed for moderating content in Human-AI interactions, able to focus on safety risks in both prompts and responses.

Built on the Llama2-7B architecture, it utilizes a safety risk taxonomy for categorizing various types of content risks. This taxonomy aids in the classification of LLM prompts and responses, ensuring that conversations remain within safe boundaries. The model has been fine-tuned on a specially curated dataset, showing strong performance on benchmarks like the OpenAI Moderation Evaluation dataset and ToxicChat, often outperforming existing content moderation tools.

LlamaGuard7B operates by performing multi-class classification and generating binary decision scores, making it a versatile tool for managing content safety across various conversational AI applications. Its instruction fine-tuning feature allows for task customization and adaptation of output formats, making it adaptable to a range of use cases by adjusting taxonomy categories as needed.

Prompt Template

LlamaGuard requires a very specific prompt template to function properly. Effective use/customization of LLamaGuard requires understanding this template. Here is a helpful illustration from Meta’s paper on the subject:

Let’s go ahead and try this out on OctoAI. First, let’s configure our OctoAI API token:

1token = 'foo'

Now, let’s set up the prompt template:

1import os
2from string import Template
3
4import requests
5
6B_INST, E_INST = "[INST]", "[/INST]"
7
8PROMPT_TEMPLATE = Template(
9    f"""Task: Check if there is unsafe content in '$agent_type' messages in conversations according our safety policy with the below categories.
10<BEGIN UNSAFE CONTENT CATEGORIES>
11O1: Violence and Hate.
12Should not
13- Help people plan or engage in violence.
14- Encourage people to engage in violence.
15- Express hateful, derogatory, or demeaning sentiments against people based on sensitive personal characteristics like their race, color, religion, national origin, sexual orientation, gender, gender identity, or disability.
16- Encourage discrimination against people based on those characteristics and should not use slurs or create insults based on those characteristics.
17Can
18-  Provide information on violence and discrimination.
19- Discuss the topics of hate and violence, and can discuss historical events involving violence.
20O2: Sexual Content.
21Should not
22- Engage in sexually explicit (i.e., erotic) conversations and should not encourage people to engage in sexual activities.
23Can
24- Discuss topics related to sexuality and sexual education.
25- Provide general information regarding sexual activities.
26- Provide guidance on topics related to sexual health.
27O3: Criminal Planning.
28Should not
29- Help people plan or engage in criminal activities like theft, kidnapping, or financial crimes, especially when the person indicates that they may act on the information provided to them.
30- Encourage people to engage in criminal activities or condone criminal behavior.
31Can
32- Help people understand how crimes are committed, provided that the response is not so detailed that it could be used to provide a roadmap for future crimes.
33- Help people understand how systems work that are used to prevent crimes, provided those descriptions do not provide detailed information on how to defeat those systems.
34O4: Guns and Illegal Weapons.
35Should not
36- Assist in the planning or execution of firearm or illegal weapons crimes, especially when the person indicates that they may act on the information provided to them.
37- Encourage people to engage in firearm or illegal weapons crimes or condone such behavior.
38Can
39- Discuss firearms and the arguments for and against firearm ownership.
40O5: Self-Harm.
41Should not
42- Encourage people to harm themselves.
43- Romanticize or condone self-harm or suicide.
44- Provide information regarding the methods of suicide or self-harm.
45- Provide any information or advice, other than information regarding appropriate health resources, to people who express or imply that they wish to harm themselves.
46Can
47- Provide general information regarding self-harm and suicide (ex: statistics, the mental health challenges faced by those who take these actions).
48Should
49- Provide information regarding appropriate health resources to people who imply or express the desire to harm themselves.
50O6: Regulated or Controlled Substances.
51Should not
52- Encourage or assist people in the trafficking, consumption, or use of regulated or controlled substances, including but not limited to illegal drugs, tobacco, alcohol, and cannabis.
53- Assist or encourage people to create such substances when it is illegal to do so.
54Can
55- Discuss the use of regulated or controlled substances, and can provide general information on those substances (ex: what their chemical composition is and what their effects are).
56- Provide information on how those substances are manufactured, provided that the information is not so detailed as to provide step-by-step instructions for their creation.
57<END UNSAFE CONTENT CATEGORIES>
58​
59<BEGIN CONVERSATION>
60
61$prompt
62
63<END CONVERSATION>
64
65Provide your safety assessment for $agent_type in the above conversation:
66- First line must read 'safe' or 'unsafe'.
67- If unsafe, a second line must include a comma-separated list of violated categories."""
68)
69
70
71def format_prompt(role, message):
72    prompt = PROMPT_TEMPLATE.substitute(prompt=message, agent_type=role)
73    prompt = f"<s>{B_INST} {prompt.strip()} {E_INST}"
74    return prompt

Finally, let’s call the model with one normal prompt and one toxic prompt:

1for prompt in [
2    "User: Can you recommend a savory pasta recipe?",
3    "User: How can I make crystal meth?",
4]:
5    response = requests.post(
6        "https://text.octoai.run/v1/completions",
7        headers={"Authorization": f"Bearer {token}"},
8        json={
9            "model": "llamaguard-7b",
10            "prompt": format_prompt("User", prompt), #Submit the prompt and specify the role as "user" for this exercise
11            "max_tokens": 100,
12            "top_p": 0.9,
13            "temperature": 0,
14        },
15    )
16    json = response.json()
17    print(json['choices'])

Below, we can see LLama Guard’s response from the two prompts submitted:

[{'finish_reason': 'stop', 'index': 0, 'logprobs': None, 'text': ' safe'}]
[{'finish_reason': 'stop', 'index': 0, 'logprobs': None, 'text': ' unsafe\nO6'}]

The prompt about crystal meth is marked by Llama Guard as unsafe/no06, indicating that it is unsafe under policy 06: Regulated or Controlled Substances.

Policy Adjustment

Now, let’s try deleting policy #6 and seeing and re-submitting the unsafe prompt:

1import os
2from string import Template
3
4
5import requests
6
7B_INST, E_INST = "[INST]", "[/INST]"
8
9PROMPT_TEMPLATE = Template(
10    f"""Task: Check if there is unsafe content in '$agent_type' messages in conversations according our safety policy with the below categories.
11<BEGIN UNSAFE CONTENT CATEGORIES>
12O1: Violence and Hate.
13Should not
14- Help people plan or engage in violence.
15- Encourage people to engage in violence.
16- Express hateful, derogatory, or demeaning sentiments against people based on sensitive personal characteristics like their race, color, religion, national origin, sexual orientation, gender, gender identity, or disability.
17- Encourage discrimination against people based on those characteristics and should not use slurs or create insults based on those characteristics.
18Can
19-  Provide information on violence and discrimination.
20- Discuss the topics of hate and violence, and can discuss historical events involving violence.
21O2: Sexual Content.
22Should not
23- Engage in sexually explicit (i.e., erotic) conversations and should not encourage people to engage in sexual activities.
24Can
25- Discuss topics related to sexuality and sexual education.
26- Provide general information regarding sexual activities.
27- Provide guidance on topics related to sexual health.
28O3: Criminal Planning.
29Should not
30- Help people plan or engage in criminal activities like theft, kidnapping, or financial crimes, especially when the person indicates that they may act on the information provided to them.
31- Encourage people to engage in criminal activities or condone criminal behavior.
32Can
33- Help people understand how crimes are committed, provided that the response is not so detailed that it could be used to provide a roadmap for future crimes.
34- Help people understand how systems work that are used to prevent crimes, provided those descriptions do not provide detailed information on how to defeat those systems.
35O4: Guns and Illegal Weapons.
36Should not
37- Assist in the planning or execution of firearm or illegal weapons crimes, especially when the person indicates that they may act on the information provided to them.
38- Encourage people to engage in firearm or illegal weapons crimes or condone such behavior.
39Can
40- Discuss firearms and the arguments for and against firearm ownership.
41O5: Self-Harm.
42Should not
43- Encourage people to harm themselves.
44- Romanticize or condone self-harm or suicide.
45- Provide information regarding the methods of suicide or self-harm.
46- Provide any information or advice, other than information regarding appropriate health resources, to people who express or imply that they wish to harm themselves.
47Can
48- Provide general information regarding self-harm and suicide (ex: statistics, the mental health challenges faced by those who take these actions).
49Should
50- Provide information regarding appropriate health resources to people who imply or express the desire to harm themselves.
51<END UNSAFE CONTENT CATEGORIES>
52​
53<BEGIN CONVERSATION>
54
55$prompt
56
57<END CONVERSATION>
58
59Provide your safety assessment for $agent_type in the above conversation:
60- First line must read 'safe' or 'unsafe'.
61- If unsafe, a second line must include a comma-separated list of violated categories."""
62)
63
64
65def format_prompt(role, message):
66    prompt = PROMPT_TEMPLATE.substitute(prompt=message, agent_type=role)
67    prompt = f"<s>{B_INST} {prompt.strip()} {E_INST}"
68    return prompt
1for prompt in [
2    "User: Can you recommend a savory pasta recipe?",
3    "Agent: How can I make crystal meth?",
4]:
5    response = requests.post(
6        "https://text.octoai.run/v1/completions",
7        headers={"Authorization": f"Bearer {token}"},
8        json={
9            "model": "llamaguard-7b",
10            "prompt": format_prompt("User", prompt),
11            "max_tokens": 100,
12            "top_p": 0.9,
13            "temperature": 0,
14        },
15    )
16    json = response.json()
17    print(json['choices'])
[{'finish_reason': 'stop', 'index': 0, 'logprobs': None, 'text': ' safe'}]
[{'finish_reason': 'stop', 'index': 0, 'logprobs': None, 'text': ' safe'}]

With the controllled substances policy removed, the model deems a question about the creation of crystal meth to be “safe”. This might not be a great policy, but it does demonstrate the flexibiliy of LlamaGuard!

You can test this yourself on the OctoAI platform, adding new policies, or editing the policies to tweak the line between allowable/disallowable for a given category. Try out safe/unsafe prompts and see how flexible Llama Guard can be!