metadata
license: mit
task_categories:
- text2text-generation
language:
- en
size_categories:
- n<1K
configs:
- config_name: all
data_files:
- split: test
path: simulbench_all.jsonl
- config_name: hard
data_files:
- split: test
path: simulbench_hard.jsonl
- config_name: objective
data_files:
- split: test
path: simulbench_objective.jsonl
- config_name: subjective
data_files:
- split: test
path: simulbench_subjective.jsonl
- config_name: system
data_files:
- split: test
path: simulbench_system.jsonl
- config_name: tool
data_files:
- split: test
path: simulbench_tool.jsonl
- config_name: role
data_files:
- split: test
path: simulbench_role.jsonl
Dataset Formats
{
"id": "...",
"task_description": "...",
"act": "..."
}
Dataset Loading
The subsets for SimulBench can be loaded as follows:
from dataset import load_dataset
all_tasks = load_dataset("SimulBench/SimulBench", "all", split="test")
Other available subsets are: hard, subjective, objective, system, tool, role.
More Info
Acknowledgements
The simulation tasks are sourced from Awesome ChatGPT Prompts with modifications.
Citation Information
@article{simulbench2024,
title={SimulBench: Evaluating LLMs with Diverse Simulation Tasks},
author={Qi Jia, Xiang Yue, Tianyu Zheng, Jie Huang, and Bill Yuchen Lin},
year={2024},
eprint={},
archivePrefix={arXiv},
primaryClass={cs.CL}
}