File size: 6,917 Bytes
07cf967
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
---

license: openrail
task_categories:
  - table-question-answering
tags:
  - power-grid
  - material-management
  - table-reasoning
  - TableQA
  - synthetic
  - domain-specific
pretty_name: Power-TableQA (Synthetic Demonstration Subset)
size_categories:
  - 1K<n<10K
configs:
  - config_name: high_complexity
    data_files:
      - split: train
        path: "high complexity/train/*"
      - split: validation
        path: "high complexity/validation/*"
      - split: test
        path: "high complexity/test/*"
  - config_name: low_complexity
    data_files:
      - split: train
        path: "low complexity/train/*"
      - split: validation
        path: "low complexity/validation/*"
      - split: test
        path: "low complexity/test/*"
---


# Power-TableQA: Synthetic Demonstration Subset

## Dataset Description

**Power-TableQA** is a domain-specific Table Question Answering (TableQA) dataset for the power grid material management domain. This repository provides a **synthetic demonstration subset** constructed using our open-sourced dataset construction pipeline.

> **Note on Data Privacy:** The full Power-TableQA dataset is built upon real operational data from Shenzhen Power Supply Co., Ltd. and cannot be publicly released due to the *Data Security Law of the People's Republic of China* and internal enterprise compliance regulations. This synthetic subset is generated entirely from simulated, non-proprietary inputs using the same construction methodology, with all sensitive fields removed or anonymized. It faithfully demonstrates the dataset structure and question types without involving any real business data.

This dataset accompanies the paper:

> **A Prior-Guided Adaptive Framework for Table Reasoning in Power Grid's Material Management**

---

## Dataset Structure

The dataset is provided as **tar archives** (one per difficulty level), each containing a Hugging Face `DatasetDict` in Arrow format with three standard splits: `train`, `validation`, and `test`.

### Configurations

| Configuration | Description |
|---|---|
| `high_complexity` | QA pairs requiring multi-step reasoning, numerical aggregation, and cross-row logical inference |
| `low_complexity` | QA pairs requiring single-hop factual lookup and straightforward value retrieval |

### Splits

| Split | Purpose |
|---|---|
| `train` | Training set |
| `validation` | Development / hyperparameter tuning set |
| `test` | Held-out evaluation set |

---

## Data Fields

Each example contains the following fields:

| Field | Type | Description |
|---|---|---|
| `feta_id` | `int64` | Unique identifier for the QA instance |
| `table_source_json` | `string` | Full table content serialized as a JSON string |
| `page_wikipedia_url` | `string` | Source document reference (simulated URL placeholder for synthetic data) |
| `table_page_title` | `string` | Title of the document or report from which the table is extracted |
| `table_section_title` | `string` | Section heading under which the table appears |
| `table_array` | `List[List[string]]` | 2-D array representation of the table (rows × columns), including the header row |
| `highlighted_cell_ids` | `List[List[int]]` | Cell coordinates `[row_idx, col_idx]` relevant to answering the question |
| `question` | `string` | Natural language question posed over the table |
| `answer` | `string` | Ground-truth natural language answer |

---

## Dataset Creation

### Construction Pipeline

The synthetic subset is generated using our **one-generation-two-verification** methodology (Section 4.1 of the paper):

1. **Table Construction** — Domain-specific tables are synthesized in the format of material management records (e.g., procurement plans, inventory schedules, supplier summaries), mimicking the schema of real power grid operational tables while using entirely simulated values.
2. **QA Generation** — Questions and answers are generated via **Template A** (Appendix A of the paper) using a large language model, covering three question types: factual lookup, numerical aggregation, and logical reasoning.
3. **QA Verification** — Generated QA pairs undergo a two-stage LLM-based verification via **Template B** (Appendix B of the paper) to ensure correctness, answerability, and alignment with the table content.

### Sensitive Field Removal

To comply with enterprise data security requirements, the following categories of sensitive information were removed or replaced with synthetic placeholders before data synthesis:

- Supplier names and identifiers
- Procurement contract references and volumes
- Internal scheduling codes and department identifiers
- Any field whose schema distribution could constitute sensitive business intelligence

---

## Usage

### Loading via `datasets`

```python

from datasets import load_dataset



# Load high complexity configuration

dataset = load_dataset("https://huggingface.co/datasets/lianqinghikari/Power-TableQA", name="high_complexity")



# Access splits

train_data = dataset["train"]

val_data   = dataset["validation"]

test_data  = dataset["test"]



# Inspect a sample

print(train_data[0]["question"])

print(train_data[0]["answer"])

print(train_data[0]["table_array"])   # 2-D table as list of lists

```

### Loading from a Local tar Archive

```python

from datasets import load_from_disk



# Extract the tar archive first, then load

dataset = load_from_disk("path/to/high complexity")

print(dataset)

```

### Example Record

```json

{

  "feta_id": 1001,

  "table_page_title": "Q3 Material Procurement Summary",

  "table_section_title": "Electrical Cable Inventory",

  "table_array": [

    ["Material Code", "Material Name", "Unit", "Planned Qty", "Actual Qty", "Difference"],

    ["M-0021", "High-voltage Cable 10kV", "m", "500", "480", "-20"],

    ["M-0034", "Low-voltage Cable 0.4kV", "m", "300", "315", "+15"],

    ["M-0087", "Transformer Bushing", "pcs", "50", "50", "0"]

  ],

  "highlighted_cell_ids": [[1, 5], [2, 5]],

  "question": "Which materials had a shortfall between planned and actual procurement quantities?",

  "answer": "High-voltage Cable 10kV had a shortfall of 20 meters."

}

```

> *For data privacy considerations, all data in this dataset is simulated and does not correspond to real operational records.*

---

## Question Types

| Type | Description | Example |
|---|---|---|
| **Factual Lookup** | Retrieve a specific cell value from the table | "What is the unit price of item M-0021?" |
| **Numerical Aggregation** | Compute sum, average, max, min across rows | "What is the total planned procurement quantity across all materials?" |
| **Logical Reasoning** | Multi-hop inference, comparison, or conditional lookup | "Which materials exceeded their planned quantity, and by how much?" |

---