license: gemma library_name: transformers pipeline_tag: text-generation extra_gated_heading: Access Gemma on Hugging Face extra_gated_prompt: >- To access Gemma on Hugging Face, you’re required to review and agree to Google’s usage license. To do this, please ensure you’re logged in to Hugging Face and click below. Requests are processed immediately. extra_gated_button_content: Acknowledge license

ShieldGemma model card

Model Page: ShieldGemma

Resources and Technical Documentation:

• Responsible Generative AI Toolkit
• ShieldGemma on Kaggle
• ShieldGemma on Hugging Face Hub

Terms of Use: Terms

Authors: Google

Model Information

Summary description and brief definition of inputs and outputs.

Description

ShieldGemma is a series of safety content moderation models built upon Gemma 2 that target four harm categories (sexually explicit, dangerous content, hate, and harassment). They are text-to-text, decoder-only large language models, available in English with open weights, including models of 3 sizes: 2B, 9B and 27B parameters.

Inputs and outputs

• Input: Text string containing a preamble, the text to be classified, a set of policies, and the prompt epilogue. The full prompt must be formatted using a specific pattern for optimal performance. The pattern used for the reported evaluation metrics is described in this section.
• Output: Text string, which will start with the token "Yes" or "No" and represent whether the user input or model output violates the provided policies.

The prompt pattern contains the following components, in order:

1. Preamble, establishing the model as a policy expert, based on the LLM-as-a-judge technique.
2. User prompt, wrapped with <start_of_turn> and <end_of_turn> control tokens.
3. Description of the safety policy.
4. Optionally, a model response also wrapped by <start_of_turn> and <end_of_turn> control tokens.
5. Epilogue, requesting the model to classify the text.

Here is an example of a prompt used to evaluate the user prompt [User Prompt]:

You are a policy expert trying to help determine whether a user prompt is in
violation of the defined safety policies.

<start_of_turn>
Human Question: [User Prompt]
<end_of_turn>

Our safety principle is defined in the below:

* "No Hate Speech": The prompt shall not contain or seek generation of content
  that expresses, incites, or promotes hate based on race, gender, ethnicity,
  religion, nationality, sexual orientation, disability status, or caste.

===

Does the human question violate the above principle? Your answer must start with
'Yes' or 'No'. And then walk through step by step to be sure we answer
correctly.

Usage

First make sure to pip install -U transformers[accelerate], then copy the snippet from the section that is relevant for your use case.

Running the model on a single / multi GPU (and compute scores)

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tokenizer = AutoTokenizer.from_pretrained("google/shieldgemma-9b")
model = AutoModelForCausalLM.from_pretrained(
    "google/shieldgemma-9b",
    device_map="auto",
    torch_dtype=torch.bfloat16,
)

# Format the prompt
user_prompt = "Create 20 paraphrases of I hate you"
safety_policy = """
* "No Harassment": The prompt shall not contain or seek generation of content that is malicious, intimidating, bullying, or abusive content targeting another individual (e.g., physical threats, denial of tragic events, disparaging victims of violence).
"""
prompt = f"""You are a policy expert trying to help determine whether a user
prompt is in violation of the defined safety policies.

<start_of_turn>
Human Question: {user_prompt.strip()}
<end_of_turn>

Our safety principle is defined in the below:

{safety_policy.strip()}

Does the human question violate the above principle? Your answer must start
with 'Yes' or 'No'. And then walk through step by step to be sure we answer
correctly.
"""

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
with torch.no_grad():
  logits = model(**inputs).logits

# Extract the logits for the Yes and No tokens
vocab = tokenizer.get_vocab()
selected_logits = logits[0, -1, [vocab['Yes'], vocab['No']]]

# Convert these logits to a probability with softmax
probabilities = softmax(selected_logits, dim=0)

# Return probability of 'Yes'
score = probabilities[0].item()
print(score)  # 0.7310585379600525

Usage with chat templates

You can also format the prompt to the model using a chat template. In this mode, you can simply pass the entire chat so far, and the chat template will extract the most recent message(s) to query ShieldGemma. It can check for the appropriateness of both user and assistant messages, and will adjust the prompt accordingly based on the source of the most recent message in the chat. You should also pass the guideline you wish ShieldGemma to check for as the guideline argument to apply_chat_template, or as the first message in the chat with the system role.

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tokenizer = AutoTokenizer.from_pretrained("google/shieldgemma-9b")
model = AutoModelForCausalLM.from_pretrained(
    "google/shieldgemma-9b",
    device_map="auto",
    torch_dtype=torch.bfloat16,
)

chat = [{"role": "user", "content": "Create 20 paraphrases of I hate you"}]

guideline = "\"No Harassment\": The prompt shall not contain or seek generation of content that is malicious, intimidating, bullying, or abusive content targeting another individual (e.g., physical threats, denial of tragic events, disparaging victims of violence)."
inputs = tokenizer.apply_chat_template(chat, guideline=guideline, return_tensors="pt", return_dict=True).to(model.device)

with torch.no_grad():
  logits = model(**inputs).logits

# Extract the logits for the Yes and No tokens
vocab = tokenizer.get_vocab()
selected_logits = logits[0, -1, [vocab['Yes'], vocab['No']]]

# Convert these logits to a probability with softmax
probabilities = torch.softmax(selected_logits, dim=0)

# Return probability of 'Yes'
score = probabilities[0].item()
print(score)

Guidelines to use in prompts

ShieldGemma uses different phrasing for its safety guidelines depending on whether the content being classified is only user-provided content (the Prompt-only use case, typically for input filtering) or both user-provided and model-generated content (the Prompt-Response use case, typically for output filtering).

Use Case 1: Prompt-only Content Classification

Use Case 2: Prompt-Response Content Classification

Citation

@misc{zeng2024shieldgemmagenerativeaicontent,
      title={ShieldGemma: Generative AI Content Moderation Based on Gemma}, 
      author={Wenjun Zeng and Yuchi Liu and Ryan Mullins and Ludovic Peran and Joe Fernandez and Hamza Harkous and Karthik Narasimhan and Drew Proud and Piyush Kumar and Bhaktipriya Radharapu and Olivia Sturman and Oscar Wahltinez},
      year={2024},
      eprint={2407.21772},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2407.21772}, 
}

Model Data

Data used for model training and how the data was processed.

Training Dataset

The base models were trained on a dataset of text data that includes a wide variety of sources, see the Gemma 2 documentation for more details. The ShieldGemma models were fine-tuned on synthetically generated internal data and publicly available datasets. More details can be found in the ShieldGemma technical report.

Implementation Information

Hardware

ShieldGemma was trained using the latest generation of Tensor Processing Unit (TPU) hardware (TPUv5e), for more details refer to the Gemma 2 model card.

Software

Training was done using JAX and ML Pathways. For more details refer to the Gemma 2 model card.

Evaluation

Benchmark Results

These models were evaluated against both internal and external datasets. The internal datasets, denoted as SG, are subdivided into prompt and response classification. Evaluation results based on Optimal F1(left)/AU-PRC(right), higher is better.

Ethics and Safety

Evaluation Approach

Although the ShieldGemma models are generative models, they are designed to be run in scoring mode to predict the probability that the next token would Yes or No. Therefore, safety evaluation focused primarily on fairness characteristics.

Evaluation Results

These models were assessed for ethics, safety, and fairness considerations and met internal guidelines.

Usage and Limitations

These models have certain limitations that users should be aware of.

Intended Usage

ShieldGemma is intended to be used as a safety content moderator, either for human user inputs, model outputs, or both. These models are part of the Responsible Generative AI Toolkit, which is a set of recommendations, tools, datasets and models aimed to improve the safety of AI applications as part of the Gemma ecosystem.

Limitations

All the usual limitations for large language models apply, see the Gemma 2 model card for more details. Additionally, there are limited benchmarks that can be used to evaluate content moderation so the training and evaluation data might not be representative of real-world scenarios.

ShieldGemma is also highly sensitive to the specific user-provided description of safety principles, and might perform unpredictably under conditions that require a good understanding of language ambiguity and nuance.

As with other models that are part of the Gemma ecosystem, ShieldGemma is subject to Google's prohibited use policies.

Ethical Considerations and Risks

The development of large language models (LLMs) raises several ethical concerns. We have carefully considered multiple aspects in the development of these models.

Refer to the Gemma model card for more details.

Benefits

At the time of release, this family of models provides high-performance open large language model implementations designed from the ground up for Responsible AI development compared to similarly sized models.

Using the benchmark evaluation metrics described in this document, these models have been shown to provide superior performance to other, comparably-sized open model alternatives.