ONESTRUCTION/Ishigaki-AEC-9B

Ishigaki-AEC-9B

Ishigaki-AEC-9B is a language model for question answering about Japanese construction specifications. Developed by ONESTRUCTION through supervised fine-tuning (SFT) of Qwen3.5-9B using LoRA (Low-Rank Adaptation), it covers architectural, civil engineering, electrical, and mechanical works.

The model is trained to answer questions in Japanese and cite the clauses that support its answers. Its training data were constructed from Japanese public construction specifications and related laws, including specifications published by Japan’s Ministry of Land, Infrastructure, Transport and Tourism (MLIT) and law texts available through e-Gov.

Use Cases

Use Ishigaki-AEC-9B to ask about construction requirements, clarify specified procedures, and identify supporting clauses during specification review.

For example, a user can ask:

公共建築工事標準仕様書において、受注者は隣接工事の請負業者とどのように協力する必要がありますか。

The model is designed to explain the requirement and provide a clause reference. Answers and cited clauses should be checked against the specification edition applicable to the project.

Supported Languages and Scope

Language: Japanese

Construction domains:

• Architectural works
• Civil engineering
• Electrical works
• Mechanical works

Source specifications and laws:

• 公共建築工事標準仕様書 — 建築工事編・電気設備工事編・機械設備工事編
• 土木工事共通仕様書 — 第1編 共通編・第2編 材料編・第3編 土木工事共通編
• Related construction laws, including 建築基準法, available through e-Gov

The source specifications are published by MLIT and its regional development bureaus. ONESTRUCTION processed the source materials to construct training and evaluation data. Applicable source terms are available from MLIT, the Kinki Regional Development Bureau, and the Kanto Regional Development Bureau.

Benchmark Score

We evaluated Ishigaki-AEC-9B on an internal benchmark consisting of 981 four-choice questions, constructed from held-out source passages that were not used in our fine-tuning data. The model achieved a score of 56.07%, compared with 43.32% for Qwen3.5-9B, an improvement of 12.75 percentage points.

The following figure compares benchmark scores and inference time per question for Ishigaki-AEC-9B and the reference models.

Usage

The example below uses Hugging Face Transformers with Qwen3.5 support. Set enable_thinking=False when applying the chat template; omitting it has been observed to cause degraded responses or unintended role tags for this checkpoint.

import torch
from transformers import AutoTokenizer, Qwen3_5ForConditionalGeneration

model_name = "ONESTRUCTION/Ishigaki-AEC-9B"
SYSTEM_PROMPT = (
    "You are a helpful assistant with specialized knowledge of Japanese public "
    "construction and civil engineering standard specifications and related laws."
)

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = Qwen3_5ForConditionalGeneration.from_pretrained(
    model_name,
    dtype="auto",
    device_map="auto",
).eval()

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {
        "role": "user",
        "content": (
            "公共建築工事標準仕様書において、受注者は隣接工事の請負業者と"
            "どのように協力する必要がありますか。"
        ),
    },
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)
inputs = tokenizer(
    text, return_tensors="pt", add_special_tokens=False
).to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=1024)

response_ids = outputs[0, inputs.input_ids.shape[1]:]
print(tokenizer.decode(response_ids, skip_special_tokens=True))

Developers

• Koyo Hidaka
• Ryo Kanazawa
• Tomoki Ando

In alphabetical order

License

Creative Commons Attribution 4.0 International (CC BY 4.0)

The base model's Apache 2.0 license also applies.

Citation

@misc{ishigaki_aec_9b_2026,
  title={{Ishigaki-AEC-9B}},
  author={{ONESTRUCTION Inc.}},
  year={2026},
  url={https://huggingface.co/ONESTRUCTION/Ishigaki-AEC-9B}
}