Datasets:

callmephilip
/

LongDA

	---
	license: cc-by-4.0
	pretty_name: LongDA
	task_categories:
	- question-answering
	- table-question-answering
	language:
	- en
	dataset_type: benchmark
	configs:
	- config_name: default
	data_files:
	- split: test
	path: benchmark.csv
	---

	[![Paper](https://img.shields.io/badge/arXiv-2601.02598-b31b1b.svg)](https://arxiv.org/pdf/2601.02598)
	[![Code](https://img.shields.io/badge/GitHub-Code-black.svg)](https://github.com/Yiyang-Ian-Li/LongDA)
	[![License](https://img.shields.io/badge/License-CC%20BY%204.0-blue.svg)](LICENSE)

	# LongDA Dataset Card

	## Dataset Description

	LongDA is a data analysis benchmark for evaluating LLM-based agents under documentation-intensive analytical workflows. It features authentic U.S. government survey data with complete, long documentation, testing LLMs' ability to navigate complex real-world datasets before performing analysis.

	### Dataset Summary

	- 505 queries extracted from 30 expert-written publications
	- 17 U.S. national surveys covering health, labor, economics, education, and social sciences
	- ~263K tokens average context per query (substantially longer than existing benchmarks)
	- Real-world analytical tasks requiring multi-step reasoning and code execution

	### Key Features

	1. Long Documentation: Multiple unstructured documents per survey (codebooks, technical reports, user guides)
	2. Expert-Grounded Queries: All queries extracted from real publications by domain experts
	3. Complex Data: Large-scale tabular data with thousands of columns requiring careful navigation
	4. Authentic Workflow: Mirrors real analytical practice where documentation navigation is the primary bottleneck

	### Languages

	- Code: Python
	- Documentation: English
	- Data: Numeric and categorical U.S. survey data

	## Dataset Structure

	⚠️ Important: To use this benchmark, you must download the complete dataset including all survey data files and documentation. The `benchmark.csv` file alone is insufficient as it only contains queries and ground truth answers. The actual evaluation requires access to raw data files and documentation in each survey folder.

	### Download Instructions

	```bash
	# Install Git LFS first
	git lfs install

	# Clone the complete dataset (recommended)
	git clone https://huggingface.co/datasets/EvilBench/LongDA benchmark

	# The dataset structure will be:
	# benchmark/
	# ├── benchmark.csv # 505 queries with ground truth answers
	# └── [SURVEY]/ # 17 survey folders (NHANES, CPS-ASEC, GSS, etc.)
	# ├── data/ # Raw survey data files (CSV, DAT, XPT, etc.)
	# └── docs/ # Long documentation (PDF, TXT)
	```

	Alternatively, download programmatically:
	```python
	from huggingface_hub import snapshot_download

	# Download entire dataset to local directory
	snapshot_download(
	repo_id="EvilBench/LongDA",
	repo_type="dataset",
	local_dir="./LongDA_data"
	)
	```


	### Data Fields

	benchmark.csv contains:
	- `survey`: Survey acronym (e.g., NHANES, CPS-ASEC)
	- `source`: Source publication title
	- `internal_id`: Question number within the publication
	- `query`: Natural language analytical query
	- `answer_structure`: Expected format (`single_number` or list structure)
	- `additional_info`: Context, units, and special requirements
	- `answer`: Ground truth answer (verified against official publications)

	## Dataset Creation

	### Source Data

	All data comes from publicly available U.S. government surveys spanning:
	- Health: Population health, nutrition, healthcare access
	- Labor & Economics: Employment, income, time use
	- Social Sciences: Demographics, drug use, family structure
	- Science & Engineering: Workforce, research funding, facilities
	- Government Operations: State finances, employment, pensions

	### Curation Process

	1. Survey Selection: Chose 17 diverse, well-documented national surveys from 6 federal agencies
	2. Publication Collection: Gathered 30 expert-written reports and publications
	3. Query Extraction: Manually extracted 505 queries grounded in real analytical practice
	4. Ground Truth Verification: Validated all answers against official statistics
	5. Documentation Assembly: Included all relevant survey documentation (codebooks, guides, technical reports)

	## Considerations for Using the Data

	### Social Impact

	This benchmark uses real government survey data covering sensitive topics including health, income, drug use, and demographics. Users should:
	- Respect data privacy and usage policies
	- Be aware of potential biases in survey data and sampling methods
	- Use results responsibly when reporting findings
	- Consider the ethical implications of automated data analysis

	### Limitations

	- Queries focus on U.S. data and may not generalize to other contexts
	- Requires significant computational resources (long context windows ~263K tokens)
	- Some surveys have complex sampling weights and methodologies
	- Documentation navigation is challenging even for humans

	## Additional Information

	### Licensing Information

	- Code and Benchmark: MIT License
	- Survey Data: Public domain (U.S. government data)
	- Documentation: Public domain (U.S. government publications)

	### Citation

	```bibtex
	@article{li2026longda,
	title={LongDA: Benchmarking LLM Agents for Long-Document Data Analysis},
	author={Li, Yiyang and Zhang, Zheyuan and Ma, Tianyi and Wang, Zehong and Murugesan, Keerthiram and Zhang, Chuxu and Ye, Yanfang},
	journal={arXiv preprint arXiv:2601.02598},
	year={2026}
	}
	```

	### Contributions

	Dataset curated by researchers at University of Notre Dame, IBM Research, and University of Connecticut. We thank all U.S. government agencies for making these valuable datasets publicly available.

	For questions or issues, please visit: https://github.com/Yiyang-Ian-Li/LongDA