SEAR

Dataset Documentation

This directory contains context-enhanced table datasets used for temporal table reasoning research presented in “No Universal Prompt: Unifying Reasoning through Adaptive Prompting for Temporal Table Reasoning”.

Overview

The dataset collection comprises 9 benchmark datasets that have been enhanced with improved table formatting using GPT-4-mini. Each dataset is stored in JSON format and contains question-answer pairs along with structured table data.

Dataset Files

Filename	Description	Questions	Domain
fetaqa.context.json	FeTaQA - Question answering over Wikipedia tables	1,582	General Knowledge
finqa.context.json	FinQA - Financial question answering over tables	962	Finance
hitabs.context.json	HiTabs - Hierarchical table question answering	897	Structured Data
hybridqa.context.json	HybridQA - Multi-hop QA over tables and text	1,528	Hybrid Reasoning
multi.context.json	Multi-hop reasoning over tables	1,587	Complex Reasoning
sqa.context.json	SQA - Sequential question answering	248	Sequential Reasoning
squall.context.json	SQUALL - SQL-like natural language QA	774	Structured Queries
tatqa.context.json	TAT-QA - Tabular and textual question answering	2,244	Hybrid Data
wiki.context.json	WikiTableQuestions - Wikipedia table QA	1,504	General Knowledge

Total: 11,326 questions across all datasets

Data Structure

Each JSON file contains an array of objects with the following fields:

{
  "_id": {
    "$oid": "unique_mongodb_object_id"
  },
  "q_num": 0,
  "question": "The question text",
  "table": "Raw table data in text/markdown format",
  "table_id": "source_table_identifier",
  "answer": "The answer or answer array",
  "improved_table_gpt4omini": "Enhanced table formatting with context"
}

Field Descriptions

• _id: MongoDB ObjectId for unique identification
• q_num: Sequential question number within the dataset
• question: Natural language question about the table
  • May be a string or array of strings for sequential questions
• table: Original table data
  • Format varies by source dataset (CSV, markdown, plain text)
  • May include section headers and metadata
• table_id: Source identifier linking back to original dataset
• answer: Question answer
  • Format varies: string, array, or list representation
  • May include numerical values, dates, names, or complex lists
• improved_table_gpt4omini: GPT-4-mini enhanced version
  • Reformatted as clean markdown tables
  • Includes contextual information and descriptions
  • Improved readability and structure

Dataset Characteristics

FeTaQA (Fact-based Table Question Answering)

• Source: Wikipedia tables
• Focus: Free-form natural language answers
• Questions require understanding table structure and content
• Example domains: Entertainment, sports, politics, history

FinQA (Financial Question Answering)

• Source: Financial reports and documents
• Focus: Numerical reasoning and financial metrics
• Questions often require calculations and comparisons
• Includes revenue, profit, growth rate calculations

HiTabs (Hierarchical Tables)

• Source: Complex hierarchical table structures
• Focus: Multi-level table understanding
• Questions span across table hierarchies

HybridQA

• Source: Wikipedia tables with associated text
• Focus: Reasoning over both tabular and textual information
• Requires multi-hop reasoning across modalities

Multi

• Source: Various table sources
• Focus: Multi-hop reasoning chains
• Complex questions requiring multiple reasoning steps

SQA (Sequential Question Answering)

• Source: Tables from Wikipedia
• Focus: Sequential question chains where context builds
• Questions arrays represent conversation-like sequences

SQUALL

• Source: WikiTableQuestions
• Focus: SQL-like natural language queries
• Structured query understanding

TAT-QA (Tabular and Textual QA)

• Source: Financial reports
• Focus: Hybrid reasoning over tables and surrounding text
• Most comprehensive financial reasoning dataset

WikiTableQuestions

• Source: Wikipedia tables
• Focus: General knowledge QA
• Diverse question types and table structures

Usage

Loading Data

import json

# Load a dataset
with open('dataset/fetaqa.context.json', 'r', encoding='utf-8') as f:
    data = json.load(f)

# Access individual examples
example = data[0]
question = example['question']
table = example['improved_table_gpt4omini']  # Use enhanced version
answer = example['answer']

Preprocessing Recommendations

1. Use improved_table_gpt4omini for better formatted tables
2. Parse answer formats based on dataset type
3. Handle sequential questions in SQA dataset as conversation chains
4. Consider table context and metadata when available

Data Enhancement

All tables have been enhanced using GPT-4-mini to:

• Standardize table formatting to clean markdown
• Add contextual descriptions and summaries
• Improve column headers and organization
• Clarify ambiguous table elements
• Provide additional context from surrounding text

Citation

If you use these datasets, please cite the original paper:

@article{sear2025,
  title={No Universal Prompt: Unifying Reasoning through Adaptive Prompting for Temporal Table Reasoning},
  author={[Authors]},
  journal={arXiv preprint},
  year={2025},
  url={https://arxiv.org/abs/2506.11246}
}

Original Dataset Sources

Please also cite the original dataset papers:

• FeTaQA: Nan et al., “FeTaQA: Free-form Table Question Answering” (2022)
• FinQA: Chen et al., “FinQA: A Dataset of Numerical Reasoning over Financial Data” (2021)
• HiTabs: Cheng et al., “HiTAB: A Hierarchical Table Dataset for Question Answering and Natural Language Generation” (2022)
• HybridQA: Chen et al., “HybridQA: A Dataset of Multi-Hop Question Answering over Tabular and Textual Data” (2020)
• SQA: Iyyer et al., “Search-based Neural Structured Learning for Sequential Question Answering” (2017)
• SQUALL: Shi et al., “SQUALL: Controlled Natural Language-to-SQL Translation” (2020)
• TAT-QA: Zhu et al., “TAT-QA: A Question Answering Benchmark on a Hybrid of Tabular and Textual Content in Finance” (2021)
• WikiTableQuestions: Pasupat and Liang, “Compositional Semantic Parsing on Semi-Structured Tables” (2015)

License

Please refer to the original dataset licenses. This enhanced version maintains the same licensing as the source datasets.

Contact

For questions about this dataset collection, please refer to the main project README or open an issue on the GitHub repository.

This site is open source. Improve this page.