Auto-SLURP / README.md
lorashen's picture
Update README.md
ebfc481 verified
metadata
license: apache-2.0
task_categories:
  - text-generation
language:
  - en
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path:
          - train.jsonl
      - split: test
        path: test.jsonl

Auto-SLURP: A Benchmark Dataset for Evaluating Multi-Agent Frameworks in Smart Personal Assistant

Repository for the paper Auto-SLURP: A Benchmark Dataset for Evaluating Multi-Agent Frameworks in Smart Personal Assistant

requirements

To test the multi-agent frameworks, you need to first install the framework according to the instruction of the framework.

We have tested CamelAI, Langgraph, AgentLite, and AutoGEN.

1. start simulated servers

cd server
sh run.sh

2. run the test

Some external servers require API keys. Therefore, to test and send requests to these servers, please make sure to provide the necessary API keys first.

The LLM used in the example code also requires configuration. Please make sure to specify the model name and provide the corresponding API key.

Please also remember to put the data/*.csv files into ~/data. Or you can modify the data path in test.py.

cd examples/autogen
sh run.sh

3. run evaluation

We are using gpt-4 for evaluating. Please set apikey properly. For e.g., export OPENAI_API_KEY="***".

If you want to use other models, please modify eval.py. For instance, if you want to use deepseek-v3 from deepseek API, you can change the model to "deepseek-chat", and change the api to your deepseek api. You also need to change the base_url to "https://api.deepseek.com/v1".

sh eval.sh

load datasets

If you want to load datasets, and use the data in jsonl format:

from datasets import load_dataset
dataset = load_dataset("lorashen/Auto-SLURP")

Citing

If you found the data or code useful, free to cite:

@misc{shen2025autoslurpbenchmarkdatasetevaluating,
      title={Auto-SLURP: A Benchmark Dataset for Evaluating Multi-Agent Frameworks in Smart Personal Assistant}, 
      author={Lei Shen and Xiaoyu Shen},
      year={2025},
      eprint={2504.18373},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2504.18373}, 
}