Model Details

Llama-Guard is a 7B parameter Llama 2-based input-output safeguard model. It can be used for classifying content in both LLM inputs (prompt classification) and in LLM responses (response classification). It acts as an LLM: it generates text in its output that indicates whether a given prompt or response is safe/unsafe, and if unsafe based on a policy, it also lists the violating subcategories.

In order to produce classifier scores, we look at the probability for the first token, and turn that into an “unsafe” class probability. Model users can then make binary decisions by applying a desired threshold to the probability scores.

Training and Evaluation

Training Data

We use a mix of prompts that come from the Anthropic dataset and redteaming examples that we have collected in house, in a separate process from our production redteaming. In particular, we took the prompts only from the Anthropic dataset, and generated new responses from our in-house LLaMA models, using jailbreaking techniques to elicit violating responses. We then annotated Anthropic data (prompts & responses) in house, mapping labels according to the categories identified above. Overall we have ~13K training examples.

Taxonomy of harms and Risk Guidelines

As automated content risk mitigation relies on classifiers to make decisions about content in real time, a prerequisite to building these systems is to have the following components: - A taxonomy of risks that are of interest – these become the classes of a classifier. - A risk guideline that determines where we put the line between encouraged and discouraged outputs for each risk category in the taxonomy. Together with this model, we release an open taxonomy inspired by existing open taxonomies such as those employed by Google, Microsoft and OpenAI in the hope that it can be useful to the community. This taxonomy does not necessarily reflect Meta’s own internal policies and is meant to demonstrate the value of our method to tune LLMs into classifiers that show high performance and high degrees of adaptability to different policies.

The Llama-Guard Safety Taxonomy & Risk Guidelines

Below, we provide both the harm types themselves under this taxonomy and also examples of the specific kinds of content that would be considered harmful under each category:

• Violence & Hate encompasses statements that encourage or could help people plan or engage in violence. Similarly, statements that advocate discrimination, contain slurs, or voice hateful sentiments against people based on their sensitive personal characteristics (ex: race, color, religion, national origin, sexual orientation, gender, gender identity, or disability) would also be considered inappropriate under this category.
• Sexual Content encompasses statements encouraging someone (who could be underage) to engage in specific sex acts. Similarly, sexually explicit (i.e., erotic) statements would also be considered inappropriate under this category.
• Guns & Illegal Weapons encompasses statements that specifically encourage, condone, or could help people plan or execute specific crimes involving the illegal acquisition, creation, or use of guns or illegal weapons (ex: explosives, biological agents, or chemical weapons), including instructions on how to create or use them.
• Regulated or Controlled Substances encompasses statements that specifically encourage or could help people to illegally produce, transfer, or consume regulated or controlled substances like illegal drugs, tobacco, alcohol, or cannabis.
• Suicide & Self Harm encompasses statements that encourage, condone, or enable people to harm themselves (ex: by providing instructions or information on methods of self-harm). When people express or imply an intent to harm themselves, any response other than one that provides links to appropriate health resources would be considered inappropriate (even if entirely supportive).
• Criminal Planning encompasses miscellaneous statements that encourage, condone, or could help people plan or execute specific criminal activities, like arson, kidnapping, or theft. Items should only be considered harmful under this category when they could not be read as violating any of the other harm types above (ex: statements that encourage violence should be considered violating under Violence & Hate rather than this category).

How to Use in transformers

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "meta-llama/LlamaGuard-7b"
device = "cuda"
dtype = torch.bfloat16

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=dtype, device_map=device)

def moderate(chat):
    input_ids = tokenizer.apply_chat_template(chat, return_tensors="pt").to(device)
    output = model.generate(input_ids=input_ids, max_new_tokens=100, pad_token_id=0)
    prompt_len = input_ids.shape[-1]
    return tokenizer.decode(output[0][prompt_len:], skip_special_tokens=True)

moderate([
    {"role": "user", "content": "I forgot how to kill a process in Linux, can you help?"},
    {"role": "assistant", "content": "Sure! To kill a process in Linux, you can use the kill command followed by the process ID (PID) of the process you want to terminate."},
])
# `safe`

You need to be logged in to the Hugging Face Hub to use the model.

For more details, see this Colab notebook.

Evaluation results

We compare the performance of the model against standard content moderation APIs in the industry, including OpenAI, Azure Content Safety,and PerspectiveAPI from Google on both public and in-house benchmarks. The public benchmarks include ToxicChat and OpenAI Moderation.

Note: comparisons are not exactly apples-to-apples due to mismatches in each taxonomy. The interested reader can find a more detailed discussion about this in our paper: [LINK TO PAPER].