REALM-Bench: A Benchmark for Evaluating Multi-Agent Systems on Real-world, Dynamic Planning and Scheduling Tasks
Paper • 2502.18836 • Published • 1
Error code: StreamingRowsError
Exception: ValueError
Message: Bad split: J1. Available splits: ['train']
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 99, in get_rows_or_raise
return get_rows(
^^^^^^^^^
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "/src/services/worker/src/worker/utils.py", line 61, in get_rows
ds = load_dataset(
^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/load.py", line 1705, in load_dataset
return builder_instance.as_streaming_dataset(split=split)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/builder.py", line 1117, in as_streaming_dataset
raise ValueError(f"Bad split: {split}. Available splits: {list(splits_generators)}")
ValueError: Bad split: J1. Available splits: ['train']Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
This Hugging Face dataset hosts the clean JSSP benchmark from REALM-Bench: a real-world planning benchmark for LLMs and multi-agent systems (paper).
All instances use a single unified JSON schema, organized into four tiers (J1–J4) as four files under JSSP/.
| File | Tier | Instances | Description |
|---|---|---|---|
JSSP/J1.json |
J1 | 109 | Static benchmarks (DMU, TA, ABZ, SWV, YN) |
JSSP/J2.json |
J2 | 109 | J1 instances + dynamic disruptions |
JSSP/J3.json |
J3 | 100 | Large-scale static (200 jobs × 50 machines) |
JSSP/J4.json |
J4 | 100 | Large-scale + multiple disruptions |
Total: 418 job-shop scheduling instances
Each J*.json file has this structure:
{
"tier": "J1",
"format_version": "1.0",
"description": "Static JSSP benchmarks (DMU, TA, ABZ, SWV, YN)",
"objective": "minimize_makespan",
"num_instances": 109,
"instances": [
{
"instance_id": "rcmax_50_20_9",
"tier": "J1",
"num_jobs": 50,
"num_machines": 20,
"jobs": [
[
{"operation": 1, "machine": 19, "processing_time": 64},
{"operation": 2, "machine": 16, "processing_time": 34}
]
],
"disruptions": [],
"metadata": {
"source_file": "DMU/rcmax_50_20_9.txt",
"benchmark": "DMU",
"objective": "minimize_makespan"
}
}
]
}
See JSSP/README.md for additional field notes.
huggingface_hub
from huggingface_hub import hf_hub_download
import json
path = hf_hub_download(
repo_id="GloriaGeng/REALM-Bench",
filename="JSSP/J1.json",
repo_type="dataset",
)
with open(path) as f:
data = json.load(f)
print(f"Tier {data['tier']}: {data['num_instances']} instances")
for inst in data["instances"][:3]:
print(inst["instance_id"], inst["num_jobs"], inst["num_machines"])
from huggingface_hub import hf_hub_download
import json
for tier in ("J1", "J2", "J3", "J4"):
path = hf_hub_download("GloriaGeng/REALM-Bench", f"JSSP/{tier}.json", repo_type="dataset")
with open(path) as f:
d = json.load(f)
print(tier, d["num_instances"], "instances,", d["objective"])
git clone https://huggingface.co/datasets/GloriaGeng/REALM-Bench
| Family | Typical size | Source |
|---|---|---|
| DMU | 20×15 – 50×20 | Demirkol, Mehta, Uzsoy |
| TA | 15×15 – 100×20 | Taillard |
| ABZ | 20×15 | Adams, Balas, Zawack |
| SWV | 20×10 – 50×10 | Storer, Vaccari, Van de Velde |
| YN | 20×20 | Yamada, Nakano |
If you use this dataset, please cite REALM-Bench:
@misc{geng2025realmbenchbenchmarkevaluatingmultiagent,
title={REALM-Bench: A Benchmark for Evaluating Multi-Agent Systems on Real-world, Dynamic Planning and Scheduling Tasks},
author={Longling Geng and Edward Y. Chang},
year={2025},
eprint={2502.18836},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2502.18836},
}
| Component | License |
|---|---|
| Code (REALM-Bench repository, evaluation framework) | MIT |
| Dataset (this JSSP release on Hugging Face) | CC-BY-4.0 |
Benchmark instance data may additionally follow the licensing of original sources (OR-Library, Taillard, etc.) where applicable.