_____ _ _ __ __ _ _
| __ \ (_) | | | \/ | | | | |
| |__) | __ ___ _ ___ ___| |_ | \ / | ___ ___ _ __ ___| |__ ___ | |_
| ___/ '__/ _ \| |/ _ \/ __| __| | |\/| |/ _ \ / _ \| '_ \/ __| '_ \ / _ \| __|
| | | | | (_) | | __
8000
/ (__| |_ | | | | (_) | (_) | | | \__ \ | | | (_) | |_
|_| |_| \___/| |\___|\___|\__| |_| |_|\___/ \___/|_| |_|___/_| |_|\___/ \__|
_/ |
|__/
Version 0.1.0
A simple and modular tool to evaluate and red-team any LLM application.
Moonshot is a tool designed for AI developers and security experts to evaluate and red-team any LLM/ LLM application. In this initial version, Moonshot can be used through its interative Command Line Interface, within python notebooks (example), or even seamlessly integrated into your model development workflow to to run repeatable tests.
-
Python (at least version 3.11)
-
Virtual Environment (optional) - Good to have different environments to separate Python dependencies
- Create a virtual environment:
python -m venv venv- Activate the virtual environment:
source venv/bin/activate
-
Download the source files by cloning this repository. i.e. Git clone (via SSH):
git clone git@github.com:moonshot-admin/moonshot.git -
Change directory to project's root directory:
cd moonshot -
Install the required packages:
pip install -r requirements.txt -
Verify Moonshot is installed:
cd srcThen, run moonshot:
python -m moonshot
Two modes are available on the Moonshot CLI: Command-Based Mode and Interactive Mode.
Full list of commands in Moonshot
Initialisation
======================================================================================================
interactive Run the interactive shell.
list_connect_types Get a list of available Language Model (LLM) connection types.
list_endpoints Get a list of available Language Model (LLM) endpoints.
version Get the version of the application.
Moonshot Benchmarking
======================================================================================================
add_cookbook Add a new cookbook.
add_endpoint Add a new endpoint.
add_recipe Add a new recipe.
list_cookbooks Get a list of available cookbooks.
list_recipes Get a list of available recipes.
list_results Get a list of available results.
list_runs Get a list of available runs.
resume_run Resume an interrupted run.
run_cookbook Run a cookbook.
run_recipe Run a recipe.
view_cookbook View a cookbook.
view_results View a results file.
Moonshot RedTeaming
=======================================================================================================
end_session End the current session.
list_prompt_templates List all prompt templates available.
list_sessions List all available sessions.
new_session Add a new red teaming session.
use_context_strategy Use a context strategy.
use_prompt_template Use a prompt template.
use_session Use an existing red teaming session.
Uncategorized
======================================================================================================
alias Manage aliases
edit Run a text editor and optionally open a file with it
help List available commands or provide detailed help for a specific command
history View, run, edit, save, or clear previously entered commands
macro Manage macros
quit Exit this application
run_pyscript Run a Python script file inside the console
run_script Run commands in script file that is encoded as either ASCII or UTF-8 text
set Set a settable parameter or show current settings of parameters
shell Execute a command as if at the OS prompt
shortcuts List available shortcuts
In the command-based mode, run commands by prepending python -m moonshot. For example,
- To list all the available commands:
python -m moonshot help - To list the connector types available:
python -m moonshot list_connect_types
We recommend the interactive mode for a more efficient experience, especially if you are using Moonshot to red-team.
To enter interactive mode: python -m moonshot interactive (You should see the command prompt change to moonshot > ) For example,
- To list all the available commands:
moonshot > help - To list the connector types available:
moonshot > list_connect_types
Establish and save connections to the endpoints of the LLMs that you wish to evaluate. Moonshot currently provides easy connection to: OpenAI's GPT4 & GPT3.5, GPT2 and Llama2-13b-gptq on Hugging Face, and Anthropic's Claude2.
Important
Please note that you will most likely need to supply your own API token/key to connect to the LLM endpoints.
To connect to these models, you simply need to create an endpoint configuration file under the directory data/llm-endpoints and define the following fields in that file:
- type: The python module name of LLM that you would like to connect to. (It should be any ONE of the Python modules available at
data/llm-connection-types) - name: The name of this endpoint (It should also be the name of this file)
- uri: The URI of the LLM endpoint.
- token: Your API token/key to connect to the LLM endpoint.
- max_calls_per_second: The maximum number of API calls made to the LLM endpoint per second.
- max_concurrency: The maximum number of open concurrent connections to the LLM endpoint.
- params: The parameter(s) required to be sent to the LLM endpoint. (optional)
For example, if you wish to create an endpoint configuration file to connect to Claude 2, you can create a file named my-anthropic-claude2.json in data/llm-endpoints. The contents of my-anthropic-claude2.json should look something like this:
```
{
"type": "claude2",
"name": "my-anthropic-claude2",
"uri": "<your_endpoint_url>",
"token": "<your_api_token>",
"max_calls_per_second": 100,
"max_concurrency": 100,
"params": {}
}
```
💡Quick Start: If you have an OpenAI API key, simply edit the pre-configured endpoint at my-openai-gpt35.json, and you'll be able to start evaluating or red-teaming GPT3.5.
Connecting LLMs - Commands
list_connect_types Get a list of available LLM connection types.
add_endpoint Add a new endpoint.
list_endpoints Get a list of configured LLM endpoints.You can run <command-name> --help to better understand the useage of a command.
Through analysis of the myriad of open-source benchmarking tasks, we have identified a common structure that encapsulates the essence of these tasks.
- Benchmark Datasets: Consists of the prompts to be sent to the model and the expected target. (if any)
- Scoring Mechanism: The method to score the model response.
- Pre & Post Prompts: The Prompt Template of additional content to be appended to the prompts in the benchmark dataset before sending to the LLM.
- Recipe: Consists minimally of a benchmark dataset and the scoring mechnism(s) to be used to score it. (Prompt template is optional)
- Cookbook: A curated set of recipes to run.
To run a cookbook
-
Activate Interactive Mode:
python -m moonshot interactive -
Run the help command for run_cookbook to better understand its usage.
run_cookbook --helpTo run one prompt from the cookbook 'bbq-lite-age-cookbook' on the LLM endpoint 'my-openai-gpt35', enter:
run_cookbook -n 1 "['bbq-lite-age-cookbook']" "['my-openai-gpt35']" -
Results will be displayed as a table and stored in
src/moonshot/data/results/
Running Evaluation Benchmarks - Commands
list_cookbooks Get a list of available cookbooks.
view_cookbook View contents of a cookbook.
add_cookbook Add a new cookbook.
run_cookbook Run a cookbook.
list_prompt_templates List all prompt templates available.
list_recipes Get a list of available recipes.
add_recipe Add a new recipe.
run_recipe Run a recipe.
view_results View a specific results file.
list_results Get a list of available results.
list_runs Get a list of available runs.
resume_run Resume an interrupted run.You can run <command-name> --help to better understand the useage of a command.
To send custom prompts to LLM endpoint(s), you need to first create a session. In a session, you can send prompts to multiple LLM endpoints (each endpoint will have its own chat), utilise prompt templates, and context strategies.
- Session: A group of Chat(s), depending on the number of endpoint(s) defined in the session. A Session can utilise one Prompt Template and Context Strategy at any one time. Every chat in a session will inherit the Prompt Template and Context Strategy set for the session.
- Endpoint: The URI of the LLM API, where your prompts will be sent to.
- Chat: The history of prompts and replies with a specific endpoint. Each endpoint in a session has a chat.
- Prompt Template: Additional static information that is appended to your prompt to form the final prompt. The final prompt will be sent to the endpoint.
- Context Strategy: Information to be sent with your current prompt to give a context or background to the LLM. Currently, the context strategy availabile are past prompt(s) and response(s) in the same chat. It is defined as an integer n, where n is the number of past prompts and responses to append to the current prompt.
Sending a Prompt to LLM Endpoint(s)
-
Create a session
There are 3 parameters (name,description and list of endpoint(s)) required when creating a session. In this example, we are creating a session with 1 endpoint:
new_session my_test_sess "this is a session description" '["my-openai-gpt35"]' -
Configure a prompt template for this session (optional)
-
Once you have created a session, you can specify a prompt template to use for all the chats in this session.
use_prompt_template <name of prompt template>- To see the list of prompt templates availabile:
list_prompt_templates
- To see the list of prompt templates availabile:
-
-
Configure a context strategy for this session (optional)
-
You can also specify the context strategy for the current session.
use_context_strategy <number of previous prompts>
-
-
Send the prompt to the endpoint by typing into the terminal directly (the following example has a prompt template and context strategy set to prepend to the prompt):
-
prompt template:
auto-categorisationwith the following contents:For the following statement, provide a short word or phrase as the completion of the sentence: -
context strategy:
1(1previous prompt and response used as the context for the current prompt) -
current prompt:
the capital of cambodia is -
results:
-
Tip
You can view the current prompt template and context strategy set by referring to PT and CS respectively in the CLI.
Red Teaming - Commands
new_session Add a new red teaming session.
list_sessions List all available sessions.
use_session Use an existing red teaming session.
end_session End the current session.
list_prompt_templates List all prompt templates available.
use_prompt_template Use a prompt template.
use_context_strategy Use a context strategy.You can run <command-name> --help to better understand the useage of a command.
In this Jupyter notebook, we demonstrate how you can leverage on the Moonshot library to:
- Connect to OpenAI's GPT-3.5
- Create your own recipes and cookbooks
- Run benchmarks
The Moonshot Team <our.moonshot.team@gmail.com>
| Dataset | Source | License |
|---|---|---|
| AdvGLUE | https://adversarialglue.github.io/ | Creative Commons Attribution 4.0 International |
| Analogical Similarity | https://github.com/google/BIG-bench/tree/main/bigbench/benchmark_tasks/analogical_similarity | Apache License Version 2.0, January 2004 |
| AI2 Reasoning Challenge | https://allenai.org/data/arc | Creative Commons Attribution-ShareAlike 4.0 International |
| Auto Categorization | https://github.com/google/BIG-bench/tree/main/bigbench/benchmark_tasks/auto_categorization | Apache License Version 2.0, January 2004 |
| BBQ | https://github.com/nyu-mll/BBQ | Creative Commons Attribution 4.0 International |
| CBBQ | https://github.com/YFHuangxxxx/CBBQ | Creative Commons Attribution Share Alike 4.0 International |
| Cause and Effect | https://github.com/google/BIG-bench/tree/main/bigbench/benchmark_tasks/cause_and_effect | Apache License Version 2.0, January 2004 |
| Challenging Toxicity Prompts | https://decodingtrust.github.io/ | Creative Commons Attribution ShareAlike 4.0 International |
| Contextual Parametric Knowledge | https://github.com/google-research/language | Apache License Version 2.0, January 2004 |
| CoQA | https://stanfordnlp.github.io/coqa/ | Creative Commons Attribution 4.0 International; MSR-LA; Apache License Version 2.0, January 2004 |
| CValues | https://github.com/X-PLUG/CValues/tree/main | Apache License Version 2.0, January 2004 |
| Enron Email | https://github.com/ParakweetLabs/EmailIntentDataSet | Apache License Version 2.0, January 2004 |
| GRE Reading Comprehension | https://github.com/google/BIG-bench/tree/main/bigbench/benchmark_tasks/gre_reading_comprehension | Apache License Version 2.0, January 2004 |
| GSM8K | https://github.com/openai/grade-school-math | MIT License |
| HellaSwag | https://github.com/rowanz/hellaswag | MIT License |
| MMLU | https://github.com/hendrycks/test | MIT License |
| TR9856 | https://aclanthology.org/P15-2069.pdf | Creative Commons Attribution 3.0 International |
| Winogrande | https://github.com/allenai/winogrande | Apache-2.0 License |
| WORD | https://aclanthology.org/L18-1408/ | Creative Commons Attribution 3.0 International |
| Real Toxicity Prompts | https://github.com/allenai/real-toxicity-prompts/ | Apache License Version 2.0, January 2004 |
| Stanford Question Answering Dataset (SQuAD) | https://modestyachts.github.io/squadshifts-website/ | Creative Commons Attribution 4.0 International |
| TruthfulQA | https://github.com/sylinrl/TruthfulQA | Apache License Version 2.0, January 2004 |
| UCI Adult | https://archive.ics.uci.edu/dataset/2/adult | Creative Commons Attribution 4.0 International |
Licensed under Apache Software License 2.0.