struct-text

# StructText — SEC_WikiDB & SEC_WikiDB_subset

*Dataset card for the VLDB 2025 TaDA-workshop submission “StructText: A
Synthetic Table-to-Text Approach for Benchmark Generation with
Multi-Dimensional Evaluation” (under review).*

```python
from datasets import load_dataset

# default = SEC_WikiDB_unfiltered_all
ds = load_dataset(
"ibm-research/struct-text",
trust_remote_code=True)

# a specific configuration
subset = load_dataset(
"ibm-research/struct-text",
"SEC_WikiDB_subset_unfiltered_planned",
trust_remote_code=True)
````

---

## 1 Dataset at a glance

| Family | Size (CSV files) | Split sizes (train/dev/test) | Notes |
| ----------------------- | ---------------- | ---------------------------- | ----------------------------------------------------- |
| **SEC\_WikiDB** | ≈ 1 000 | 80 % / 10 % / 10 % | Parsed from EDGAR 10-K / 10-Q filings + WikiDB tables |
| **SEC\_WikiDB\_subset** | 49 | 39 / 5 / 5 | Handy subset used in the paper |

Each split contains three *file types*:

| Suffix | Meaning |
| ---------------- | ------------------------------------------------------- |
| `_original.csv` | Raw structured data (columns + rows) |
| `*_generated_reports_*.csv` | Text generated from the table via Qwen-2-5-72B-Instruct |
| `*_report_types_*.csv` | Reference text produced by our planning module |

---

## 2 Folder layout

```
SEC_WikiDB/
├─ unfiltered/
│ ├─ train/ *_original.csv │ *_generated.csv │ *_planned.csv
│ ├─ dev/ …
│ └─ test/ …
└─ filtered/ # <- coming soon

SEC_WikiDB_subset/
├─ unfiltered/
│ ├─ train/ *_original.csv │ *_generated.csv │ *_planned.csv
│ ├─ dev/ …
│ └─ test/ …
└─ filtered/ # <- coming soon

```

The **loader** treats
`<family>_<filtered|unfiltered>_<all|original|generated|planned>`
as *configuration names*, e.g. `SEC_WikiDB_filtered_generated`.

---

## 3 Quick-start examples

```python
# full corpus, but original tables only
orig = load_dataset("ibm-research/struct-text",
"SEC_WikiDB_unfiltered_original",
trust_remote_code=True)

# data-frame reconstruction for one CSV file
import pandas as pd, io
ex = orig['test'][0]
df = pd.DataFrame(ex["rows"], columns=ex["columns"])
```

---

## 4 Dataset creation

* **WikiDB component** — Scraped via the method in Vogel et al. 2024 \[1].
* **SEC component** — Programmatic EDGAR queries (10-K/10-Q XML) → CSV.
* **Generation & planning** — Qwen-2-5-72B-Instruct + Two stage prompting for planning followed by report generation.
* **Filtering (ongoing)** — Unit-time accuracy threshold search (see paper §3.3).

---

## 5 Citation

```
@inproceedings{kashyap2025structtext,
title = {StructText: A Synthetic Table-to-Text Approach …},
author = {Satyananda Kashyap and Sola Shirai and
Nandana Mihindukulasooriya and Horst Samulowitz},
booktitle = {Proc.\ VLDB TaDA Workshop},
year = {2025},
note = {Accepted Oral}
}
```

**Sources**

1. Liane Vogel, Jan-Micha Bodensohn, Carsten Binnig.
*WikiDBs: A Large-Scale Corpus of Relational Databases from Wikidata.*
NeurIPS 2024 Datasets & Benchmarks Track.
2. *SEC EDGAR database.* [https://www.sec.gov/edgar](https://www.sec.gov/edgar)

# StructText — SEC_WikiDB 与 SEC_WikiDB_subset

*本数据集卡片对应VLDB 2025 TaDA研讨会投稿论文《StructText：面向多维度评估基准生成的合成表到文本方法》（审稿中）。*

python
from datasets import load_dataset

# default = SEC_WikiDB_unfiltered_all
ds = load_dataset(
"ibm-research/struct-text",
trust_remote_code=True)

# a specific configuration
subset = load_dataset(
"ibm-research/struct-text",
"SEC_WikiDB_subset_unfiltered_planned",
trust_remote_code=True)

---

## 1 数据集概览

| 数据集族 | CSV文件规模 | 划分比例（训练集/验证集/测试集） | 备注 |
| ----------------------- | ---------- | ---------------------------- | ----------------------------------------------------- |
| **SEC_WikiDB** | ≈ 1 000 | 80 % / 10 % / 10 % | 从EDGAR 10-K / 10-Q申报文件与WikiDB表中解析得到 |
| **SEC_WikiDB_subset** | 49 | 39 / 5 / 5 | 论文中使用的便捷子集 |

每个划分包含三类文件类型：

| 文件后缀 | 含义 |
| ---------------- | ------------------------------------------------------- |
| `_original.csv` | 原始结构化数据（含列名与行数据） |
| `*_generated_reports_*.csv` | 通过Qwen-2-5-72B-Instruct从表格生成的文本 |
| `*_report_types_*.csv` | 由我们的规划模块生成的参考文本 |

---

## 2 文件夹结构

SEC_WikiDB/
├─ unfiltered/
│ ├─ train/ *_original.csv │ *_generated.csv │ *_planned.csv
│ ├─ dev/ …
│ └─ test/ …
└─ filtered/ # <- 即将上线

SEC_WikiDB_subset/
├─ unfiltered/
│ ├─ train/ *_original.csv │ *_generated.csv │ *_planned.csv
│ ├─ dev/ …
│ └─ test/ …
└─ filtered/ # <- 即将上线

**数据集加载器**将格式为`<数据集族>_<过滤版|非过滤版>_<全量|原始数据|生成文本|规划参考>`的字符串视为配置名称，例如`SEC_WikiDB_filtered_generated`。

---

## 3 快速入门示例

python
# 完整语料库，仅保留原始表格
orig = load_dataset("ibm-research/struct-text",
"SEC_WikiDB_unfiltered_original",
trust_remote_code=True)

# 重构单个CSV文件的数据帧
import pandas as pd, io
ex = orig['test'][0]
df = pd.DataFrame(ex["rows"], columns=ex["columns"])

---

## 4 数据集构建

* **WikiDB模块** — 通过Vogel等人2024年的方法[1]爬取得到。
* **SEC模块** — 通过程序化EDGAR查询（10-K/10-Q XML文件）转换为CSV格式。
* **生成与规划模块** — 采用Qwen-2-5-72B-Instruct，结合两阶段提示：先规划再生成报告文本。
* **过滤流程（进行中）** — 单位时间准确率阈值搜索（详见论文§3.3）。

---

## 5 引用格式

@inproceedings{kashyap2025structtext,
title = {StructText: A Synthetic Table-to-Text Approach …},
author = {Satyananda Kashyap and Sola Shirai and
Nandana Mihindukulasooriya and Horst Samulowitz},
booktitle = {Proc. VLDB TaDA Workshop},
year = {2025},
note = {Accepted Oral}
}

**参考文献**

1. Liane Vogel、Jan-Micha Bodensohn、Carsten Binnig. *WikiDBs：基于Wikidata的大规模关系数据库语料库*，发表于NeurIPS 2024数据集与基准赛道。
2. *SEC EDGAR数据库*。[https://www.sec.gov/edgar](https://www.sec.gov/edgar)