- [05/2024] Sotopia was presented at ICLR 2024 as a spotlight โญ!
Sotopia is an open-ended social learning environment that allows agents to interact with each other and the environment. The environment is designed to be a platform for evaluating and faciliating social intelligence in language agents. The environment is designed to be open-ended, meaning that the environment can be easily extended to include new environments and new agents. The environment is also designed to be scalable, meaning that the environment can be easily scaled to include a large number of agents and environments.
@inproceedings{zhou2024sotopia,
title = {SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents},
author = {Zhou*, Xuhui and Zhu*, Hao and Mathur, Leena and Zhang, Ruohong and Qi, Zhengyang and Yu, Haofei and Morency, Louis-Philippe and Bisk, Yonatan and Fried, Daniel and Neubig, Graham and Sap, Maarten},
journal = {ICLR},
year = {2024},
url = {https://openreview.net/forum?id=mM7VurbA4r},
}
See documentation for more details.
If you want to try it out on Google Colab first, please check out our Colab Tutorial Series:
- Basic
- Building your own social agent (coming soon!)
This package supports Python 3.11 and above. In one line,
pip install sotopia
or pip install uv; uv pip install sotopia
.
Or from scratch, use a virtual environment, e.g. with anaconda3: conda create -n sotopia python=3.11; conda activate sotopia; curl -sSL https://install.python-poetry.org | python3
. Then, install the requirements and this package.
poetry install
OpenAI key is required to run the code. Please set the environment variable OPENAI_API_KEY
to your key. The recommend way is to add the key to the conda environment:
conda env config vars set OPENAI_API_KEY=your_key
For some experiments, TogetherAI key is required to run the code. Please set the environment variable TOGETHER_API_KEY
to your key. The recommend way is to add the key to the conda environment:
conda env config vars set TOGETHER_API_KEY=your_key
A redis-stack server is required to run the code. Please follow the instruction to start a redis-stack server or use an existing server. You can also check Q&A to initiate the redis server with the Sotopia data.
The REDIS_OM_URL
need to be set before loading and saving agents:
conda env config vars set REDIS_OM_URL="redis://user:password@host:port"
Make a folder to store the logs:
mkdir logs
You can view an episode demo with default parameters with the following:
import asyncio
from sotopia.samplers import UniformSampler
from sotopia.server import run_async_server
asyncio.run(
run_async_server(
model_dict={
"env": "gpt-4",
"agent1": "gpt-3.5-turbo",
"agent2": "gpt-3.5-turbo",
},
sampler=UniformSampler(),
)
)
or run
python examples/minimalist_demo.py
Follow the installation instruction above and then, instead of running python -m pip install -e .
, run the following commands:
python -m pip install -e ".[dev]"
mypy --install-types --non-interactive sotopia
python -m pip install pre-commit
pre-commit install
git checkout -b feature/feature-name
and PR to main
branch.
Run pytest
to make sure all tests pass (this will ensure dynamic typing passed with beartype) and mypy --strict .
to check static typing.
(You can also run pre-commit run --all-files
to run all checks)
Check the github action result to make sure all tests pass. If not, fix the errors and push again.
We use gin-config
to configure the experiments. You don't need to be an expert to use it. The basic syntax is
python <code_file.py> --gin_file <gin_file1> --gin_file <gin_file2> '--gin.PARAM1=value1' '--gin.PARAM2=value2'
The --gin_file
is used to load and compose the default configuration. The --gin.PARAM1=value1
is used to overwrite the default configuration. The later configuration will always overwrite the previous one.
Here is an example of running an experiment:
python examples/experiment_eval.py --gin_file sotopia_conf/generation_utils_conf/generate.gin --gin_file sotopia_conf/server_conf/server.gin --gin_file sotopia_conf/run_async_server_in_batch.gin '--gin.ENV_IDS=["01H7VFHPDZVVCDZR3AARA547CY"]' '--gin.AGENT1_MODEL="gpt-4"' '--gin.BATCH_SIZE=20' '--gin.PUSH_TO_DB=False' '--gin.TAG="test"'
For the complete set of parameters, please check the sotopia_conf
folder.
To run a large batch of environments, you can change the ENV_IDS
parameter in sotopia_conf/run_async_server_in_batch.gin
to a list of environment ids. When gin.ENV_IDS==[]
, all environments on the DB will be used.
To use a custom API url for agent models, you can set AGENT1_MODEL
to the following format:
"custom/{model_name}@{base_url}"
As an example, if you use vllm to expose an OpenAI API:
python -m vllm.entrypoints.openai.api_server --model facebook/opt-125m
The gin.AGENT1_MODEL
should be set to:
"custom/facebook/opt-125m@http://localhost:8000/v1"
If your API requires a token, you can set the environment variable CUSTOM_API_KEY
to your key.
As another example, if your vllm API requires a token:
python -m vllm.entrypoints.openai.api_server --model facebook/opt-125m --api_key hello
You could set the environment variable CUSTOM_API_KEY
to hello
.
After running experiments, you can go to the examples/redis_stats.ipynb
notebook to check the existing episodes (Episode Log section), as well as calculate the performance.
For the original Sotopia simulation in our paper's experiments, you can find how to get them in the Q&A section in the ./docs
folder.
You can use the following function with the **kwargs
being the properties of the AgentProfile
class. This is the same for the scenarios/environments.
class AgentProfile(JsonModel):
first_name: str = Field(index=True)
last_name: str = Field(index=True)
age: int = Field(index=True, default_factory=lambda: 0)
occupation: str = Field(index=True, default_factory=lambda: "")
gender: str = Field(index=True, default_factory=lambda: "")
gender_pronoun: str = Field(index=True, default_factory=lambda: "")
public_info: str = Field(index=True, default_factory=lambda: "")
big_five: str = Field(index=True, default_factory=lambda: "")
moral_values: list[str] = Field(index=False, default_factory=lambda: [])
schwartz_personal_values: list[str] = Field(index=False, default_factory=lambda: [])
personality_and_values: str = Field(index=True, default_factory=lambda: "")
decision_making_style: str = Field(index=True, default_factory=lambda: "")
secret: str = Field(default_factory=lambda: "")
model_id: str = Field(default_factory=lambda: "")
class EnvironmentProfile(JsonModel):
codename: str = Field(...)
source: str = Field(...)
scenario: str = Field(...)
agent_goals: list[str] = Field(...)
...
from sotopia.database.persistent_profile import AgentProfile, EnvironmentProfile
def add_agent_to_database(**kwargs: dict[str, Any]) -> None:
agent = AgentProfile(**kwargs)
agent.save()
def add_env_profile(**kwargs: dict[str, Any]) -> None:
env_profile = EnvironmentProfile(**kwargs)
env_profile.save()