REALM-Bench — Job Shop Scheduling (JSSP) Dataset

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/.

Dataset files

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

Problem tiers

• J1 — Classic static JSSP from standard OR benchmarks (minimize makespan).
• J2 — Same problems as J1 with disruptions (machine breakdown, power outage, supply delay, etc.); tests reactive replanning.
• J3 — Industry-scale static instances (larger job and machine counts).
• J4 — J3 scale with multiple simultaneous disruptions; highest difficulty tier.

Data format

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"
      }
    }
  ]
}

• jobs — Ordered operations per job; machine is 1-based; processing_time is a positive integer.
• disruptions — Empty for J1/J3; list of events for J2/J4.

See JSSP/README.md for additional field notes.

Usage

Download with 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"])

Load all tiers

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"])

Clone the dataset

git clone https://huggingface.co/datasets/GloriaGeng/REALM-Bench

Benchmark families (J1 / J2)

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

Citation

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},
}

Links

• GitHub: https://github.com/genglongling/REALM-Bench
• Paper: https://arxiv.org/abs/2502.18836
• JSSP subfolder docs: JSSP/README.md

License

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.