yongqiqng/CivilInstruct-Sample

---
license: cc-by-nc-4.0
task_categories:
- text-generation
language:
- en
tags:
- code
- structural-engineering
- openseespy
- scientific-modeling
- instruction-tuning
- reinforcement-learning
size_categories:
- 1K<n<10K
configs:
- config_name: sft
data_files:
- split: train
path: data/sft_train.parquet
- split: validation
path: data/sft_val.parquet
- config_name: rl
data_files:
- split: train
path: data/rl_train.parquet
- split: test
path: data/rl_test.parquet
---

# CivilInstruct-Sample

A 10% sample of the **CivilInstruct** dataset from the paper *Rethinking Scientific Modeling: Toward Physically Consistent and Simulation-Executable Programmatic Generation*.

This sample is released for **demonstration and reproducibility** of the AutoBM training pipeline. The full dataset (10,912 samples) will be released upon paper publication.

## Overview

CivilInstruct is a domain-specific instruction dataset for training LLMs to generate executable, physically consistent **OpenSeesPy** structural modeling code from natural language building specifications. It integrates four data sources:

| Part | Description | Full | Sample (this repo) |
|------|-------------|------|--------------------|
| Part 1 | Fine-grained supervised data (OpenSeesPy API tutorials) | 3,881 | ~388 |
| Part 2 | Parameterized generated long code data | 3,100 | ~310 |
| Part 3 | Execution error-oriented debugging CoT data | 3,500 | ~350 |
| Part 4 | Physics-informed expert data (with ground-truth periods) | 512 | ~51 |

## Subsets

### `sft` — Supervised Fine-Tuning
Used for Stage I training (instruction tuning).

| Split | Samples | Fields |
|-------|---------|--------|
| `train` | 989 | `instruction`, `response` |
| `validation` | 20 | `instruction`, `response` |

### `rl` — Reinforcement Learning (SPC-GRPO)
Used for Stage II training with Multi-Granularity Hybrid Reward (MGHR).

| Split | Samples | Fields |
|-------|---------|--------|
| `train` | 45 | `data_source`, `prompt`, `ability`, `reward_model`, `extra_info` |
| `test` | 5 | `data_source`, `prompt`, `ability`, `reward_model`, `extra_info` |

The `reward_model.ground_truth` field contains the structural fundamental period (T1) used for physics-aware reward grading.

## Usage

```python
from datasets import load_dataset

# SFT data
sft_train = load_dataset("yongqiqng/CivilInstruct-Sample", "sft", split="train")
print(sft_train[0])

# RL data
rl_train = load_dataset("yongqiqng/CivilInstruct-Sample", "rl", split="train")
print(rl_train[0])
```

## Related

- Paper: [arXiv:2602.07083](https://arxiv.org/abs/2602.07083)
- Code: [github.com/Jovanqing/AutoBM](https://github.com/Jovanqing/AutoBM)
- Model: [Jovanqing/Seed-Coder-8B-Reasoning-SFT](https://huggingface.co/Jovanqing/Seed-Coder-8B-Reasoning-SFT)
- Full dataset: *Coming soon upon paper publication*

## Citation

```bibtex
@article{jiang2026rethinking,
title={Rethinking Scientific Modeling: Toward Physically Consistent and Simulation-Executable Programmatic Generation},
author={Jiang, Yongqing and Wang, Jianze and Shen, Zhiqi and Lin, Zhenghong and Wang, Jiayuan and Yang, Yijian and Dai, Kaoshan and Luo, Haoran},
journal={arXiv preprint arXiv:2602.07083},
year={2026}
}
```

## License

This sample is released under **CC BY-NC 4.0** (non-commercial). For commercial inquiries, please contact the authors.