Agent Skill Acquisition for Large Language Models via CycleQD (ICLR 2025)

📚 [Paper] | 🤗 [Hugging Face] | 📝 [Blog]

Installation

Basic library

pip install -r requirements.txt

Task evaluator

cd evaluation/fishfarm
pip install -e .

VLLM module

pip install vllm

Evalplus task

pip install git+https://github.com/evalplus/evalplus@1895d2f6aa8895044a7cf69defc24bd57695e885

DBBench task

Run the following commands after docker installation

docker pull mysql
pip install mysql-connector-python==8.0.32 docker==6.1.2

Tips

• If you encounter errors connecting to MySQL containers, please increase the value of /proc/sys/fs/aio-max-nr (e.g., echo 1048576 | sudo tee /proc/sys/fs/aio-max-nr)
• We use requests==2.31.0 in our setup. (see this issue)

OSInteraction task

Run the following command after docker installation

python data/os_interaction/images.py build -c data/os_interaction/configs/std.yaml -r .

Celery and redis

pip install celery  
pip install redis

Training

Single worker

python3 main.py

Multiple workers

# 1. Start rabbitmq Broker
docker run -d -p 5672:5672 -v utils/rabbitmq.conf:/etc/rabbitmq/rabbitmq.conf rabbitmq
# 2. Start redis Broker
docker run -d -p 6379:6379 -v utils/redis.conf:/etc/redis/redis.conf redis redis-server /etc/redis/redis.conf
# 3. Start Worker(s)
python3 main.py -m celery.mode=worker
# 4. Start Main
python3 main.py -m celery.mode=main

Bibtex

To cite our work, you can use the following:

@article{sakana2024cycleQD,
  title={Agent Skill Acquisition for Large Language Models via CycleQD},
  author={So Kuroki and Taishi Nakamura and Takuya Akiba and Yujin Tang},
  year={2024},
  eprint={2410.14735},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2410.14735},
}