Datasets:
metadata
license: cc-by-4.0
task_categories:
- text-generation
language:
- ja
- en
tags:
- ids
- ifc
- bim
- construction
- buildingSMART
pretty_name: Ishigaki-IDS-Bench
configs:
- config_name: default
data_files:
- split: test
path: data/test.jsonl
Ishigaki-IDS-Bench
Ishigaki-IDS-Bench is a bilingual benchmark for generating buildingSMART IDS 1.0 XML from user-facing building requirements.
Data Format
Each row uses the common conversational messages format:
{
"id": "row-0001",
"messages": [
{"role": "user", "content": "..."},
{"role": "assistant", "content": "..."}
],
"language": "ja",
"input_format": "natural_language",
"turn_type": "single_turn",
"domain": "architectural",
"ifc_versions": ["IFC4"]
}
The final assistant message is the gold IDS XML. The evaluation input is messages[:-1].
Fields
id: example identifier.messages: multi-turn conversation inrole/contentformat.language:jaoren.input_format:natural_languageorcsv.turn_type:single_turnormulti_turn.domain:architectural,structural,mep, orgeneral.ifc_versions: target IFC version labels.
Citation
@misc{kanazawa2026ishigakiidsbench,
title = {Ishigaki-IDS-Bench: A Benchmark for Generating Information Delivery Specification from BIM Information Requirements},
author = {Ryo Kanazawa and Koyo Hidaka and Teppei Miyamoto and Takayuki Kato and Tomoki Ando and Chenguang Wang and Dayuan Jiang and Naofumi Fujita and Shuhei Saitoh and Atomu Kondo and Koki Arakawa and Daiho Nishioka},
year = {2026},
eprint = {2605.22079},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
doi = {10.48550/arXiv.2605.22079},
url = {https://arxiv.org/abs/2605.22079}
}