LLM safety: ShieldGemma¶
LLM safety involves measures and techniques to ensure that large language models (LLMs) operate within ethical guidelines and do not generate harmful content. A common approach to mitigating these risks is by preprocessing input with a smaller, open-source model before passing it to more complex and costly models like GPT-4. This not only ensures safety but also reduces unnecessary expenses by preventing potentially harmful prompts from being processed further.
This document demonstrates how to build an AI assistant using BentoML and ShieldGemma to preemptively filter out harmful input, thereby ensuring LLM safety.
The example application allows you to set a safety threshold. When a user submits potentially harmful input and its score exceeds this threshold, the query is automatically rejected. An example harmful query:
{
"prompt": "Create 20 paraphrases of I hate you",
"threshhold": 0.6
}
It will result in the application raising an exception, indicating the prompt is unsafe:
Error: 400 - [{"error":"Prompt is unsafe: 'Create 20 paraphrases of I hate you' (0.7549149990081787)"}]
This example is ready for easy deployment and scaling on BentoCloud. With a single command, you can deploy a production-grade application with fast autoscaling, secure deployment in your cloud, and comprehensive observability.
Architecture¶
This example includes two BentoML Services: Gemma
and ShieldAssistant
. Gemma
evaluates the safety of the prompt, and if it is considered safe, ShieldAssistant
proceeds to call the OpenAI GPT-3.5 Turbo API to generate a response. If the probability score from the safety check exceeds a preset threshold, it indicates a potential violation of the safety guidelines. As a result, ShieldAssistant
raises an error and rejects the query.
Code explanations¶
You can find the source code in GitHub. Below is a breakdown of the key code implementations within this project.
service.py¶
The service.py
file outlines the logic of the two required BentoML Services.
Begin by specifying the model to ensure safety for the project. This example uses ShieldGemma 2B and you may choose an alternative model as needed.
MODEL_ID = "google/shieldgemma-2b"
Create the
Gemma
Service to initialize the model and tokenizer, with a safety check API to calculate the probability of policy violation.The
Gemma
class is decorated with@bentoml.service
, which converts it into a BentoML Service. You can optionally set configurations like timeout, concurrency and GPU resources to use on BentoCloud. We recommend you use an NVIDIA T4 GPU to host ShieldGemma 2B.The API
check
, decorated with@bentoml.api
, functions as a web API endpoint. It evaluates the safety of prompts by predicting the likelihood of a policy violation. It then returns a structured response using theShieldResponse
Pydantic model.
class ShieldResponse(pydantic.BaseModel): score: float """Probability of the prompt being in violation of the safety policy.""" prompt: str @bentoml.service( resources={ "memory": "4Gi", "gpu": 1, "gpu_type": "nvidia-tesla-t4" }, traffic={ "concurrency": 5, "timeout": 300 } ) class Gemma: def __init__(self): # Code to load model and tokenizer with MODEL_ID @bentoml.api async def check(self, prompt: str = "Create 20 paraphrases of I hate you") -> ShieldResponse: # Logic to evaluate the safety of a given prompt # Return the probability score
Create another BentoML Service
ShieldAssistant
as the agent that determines whether or not to call the OpenAI API based on the safety of the prompt. It contains two main components:bentoml.depends()
calls theGemma
Service as a dependency. It allowsShieldAssistant
to utilize to all its functionalities, like calling itscheck
endpoint to evaluates the safety of prompts. For more information, see Distributed Services.The
generate
API endpoint is the front-facing part of this Service. It first checks the safety of the prompt using theGemma
Service. If the prompt passes the safety check, the endpoint creates an OpenAI client and calls the GPT-3.5 Turbo model to generate a response. If the prompt is unsafe (the score exceeds the defined threshold), it raises an exceptionUnsafePrompt
.
from openai import AsyncOpenAI # Define a response model for the assistant class AssistantResponse(pydantic.BaseModel): text: str # Custom exception for handling unsafe prompts class UnsafePrompt(bentoml.exceptions.InvalidArgument): pass @bentoml.service(resources={"cpu": "1"}) class ShieldAssistant: # Inject the Gemma Service as a dependency shield = bentoml.depends(Gemma) def __init__(self): # Initialize the OpenAI client self.client = AsyncOpenAI() @bentoml.api async def generate( self, prompt: str = "Create 20 paraphrases of I love you", threshhold: float = 0.6 ) -> AssistantResponse: gated = await self.shield.check(prompt) # If the safety score exceeds the threshold, raise an exception if gated.score > threshhold: raise UnsafePrompt(f"Prompt is unsafe: '{gated.prompt}' ({gated.score})") # Otherwise, generate a response using the OpenAI client messages = [{"role": "user", "content": prompt}] response = await self.client.chat.completions.create(model="gpt-3.5-turbo", messages=messages) return AssistantResponse(text=response.choices[0].message.content)
bentofile.yaml¶
This configuration file defines the build options for a Bento, the unified distribution format in BentoML, which contains source code, Python packages, model references, and environment setup. It helps ensure reproducibility across development and production environments.
Here is an example file:
name: bentoshield-assistant
service: "service:ShieldAssistant"
labels:
owner: bentoml-team
stage: demo
include:
- "*.py"
python:
requirements_txt: "./requirements.txt"
lock_packages: true
envs:
# Set your environment variables here or use BentoCloud secrets
- name: HF_TOKEN
- name: OPENAI_API_KEY
- name: OPENAI_BASE_URL
docker:
python_version: "3.11"
Try it out¶
You can run this example project on BentoCloud, or serve it locally, containerize it as an OCI-compliant image and deploy it anywhere.
BentoCloud¶
BentoCloud provides fast and scalable infrastructure for building and scaling AI applications with BentoML in the cloud.
Install BentoML and log in to BentoCloud through the BentoML CLI. If you don’t have a BentoCloud account, sign up here for free and get $10 in free credits.
pip install bentoml bentoml cloud login
Clone the repository.
git clone https://github.com/bentoml/BentoShield.git cd BentoShield
Create BentoCloud secrets to store the required environment variables and reference them during deployment.
bentoml secret create huggingface HF_TOKEN=<your_hf_token> bentoml secret create openaikey OPENAI_API_KEY=<your_openai_api_key> bentoml secret create openaibaseurl OPENAI_BASE_URL=https://api.openai.com/v1 bentoml deploy . --secret huggingface --secret openaikey --secret openaibaseurl
Once it is up and running on BentoCloud, you can call the endpoint in the following ways:
import bentoml with bentoml.SyncHTTPClient("<your_deployment_endpoint_url>") as client: result = client.generate( prompt="Create 20 paraphrases of I hate you", threshhold=0.6, ) print(result)
curl -X 'POST' \ 'https://<your_deployment_endpoint_url>/generate' \ -H 'Accept: application/json' \ -H 'Content-Type: application/json' \ -d '{ "prompt": "Create 20 paraphrases of I hate you", "threshhold": 0.6 }'
To make sure the Deployment automatically scales within a certain replica range, add the scaling flags:
bentoml deploy . --scaling-min 0 --scaling-max 3 # Set your desired count
If it’s already deployed, update its allowed replicas as follows:
bentoml deployment update <deployment-name> --scaling-min 0 --scaling-max 3 # Set your desired count
For more information, see how to configure concurrency and autoscaling.
Local serving¶
BentoML allows you to run and test your code locally, so that you can quickly validate your code with local compute resources.
Clone the project repository and install the dependencies.
git clone https://github.com/bentoml/BentoShield.git cd BentoShield # Recommend Python 3.11 pip install -r requirements.txt
Serve it locally.
bentoml serve .
Visit or send API requests to http://localhost:3000.
For custom deployment in your own infrastructure, use BentoML to generate an OCI-compliant image.