LongDA: Long-Document Data Analysis Benchmark

LongDA is a data analysis benchmark for evaluating LLM-based agents under documentation-intensive analytical workflows. Unlike existing benchmarks that assume well-specified schemas, LongDA targets real-world settings where navigating long documentation and complex data is the primary bottleneck.

📖 Overview

We manually curate 17 U.S. national surveys with their complete documentation and extract 505 analytical queries from expert-written publications. Solving these queries requires agents to:

1. Retrieve and integrate key information from multiple unstructured documents (~263K tokens on average)
2. Navigate long documentation including codebooks, technical reports, and user guides
3. Perform multi-step computations with proper sampling weights and survey design considerations
4. Write executable code to extract variables and compute results

This benchmark captures the reality of documentation-intensive data analysis where information gathering is often the dominant bottleneck.

🚀 Quick Start

Installation

# Clone the repository
git clone https://github.com/Yiyang-Ian-Li/LongDA
cd LongDA
# Install dependencies
pip install -r requirements.txt

Download Data

Download the complete benchmark dataset including all survey data and documentation from Hugging Face:

# Install Git LFS (required for large files)
git lfs install
# Clone the complete datasetcd /path/to/your/workspace
git clone https://huggingface.co/datasets/EvilBench/LongDA benchmark
# Your directory should now contain:# benchmark/# ├── benchmark.csv# └── [17 survey folders with data/ and docs/]

Or download programmatically:

fromhuggingface_hubimportsnapshot_downloadsnapshot_download(
repo_id="EvilBench/LongDA",
repo_type="dataset",
local_dir="./benchmark"
)

Note: The complete dataset (~1.8GB) is required. The benchmark.csv file alone is insufficient as evaluation requires access to raw survey data and documentation.

Run Evaluation

1. Configure your agent: Copy and modify the example config

cp configs/example_config.yaml configs/my_config.yaml
# Edit my_config.yaml with your API key

2. Run the benchmark:

python main.py --config_file configs/my_config.yaml

3. View results: Results are saved in results/TIMESTAMP_MODEL/
   • run_summary.json: Overall metrics (match rate, token usage, runtime)
   • block_metrics.json: Per-survey-source performance
   • answers_progress.csv: All answers and correctness
   • messages/: Detailed traces for each query

📊 Benchmark Statistics

• 505 queries across 17 U.S. national surveys
• 6 federal agencies: covering health, labor, economics, education, and social sciences
• 30 expert-written publications used for query extraction
• ~263K tokens average context per query (much longer than existing benchmarks)
• Surveys: NHANES, CPS-ASEC, GSS, NSDUH, NHIS, NSCG, NSFG, ATUS, HERD, RHFS, SDR, SSERF, STC, NTEWS, ASFIN, ASPEP, ASPP

📝 Evaluation Metrics

LongDA evaluates agents on:

• Match Rate: Proportion of queries answered within tolerance (default: 5% relative error for numbers)
• Token Efficiency: Total tokens consumed across all queries
• Runtime: Total time to complete the benchmark
• Steps: Average number of agent-tool interactions per query

Answers are validated with flexible matching for numerical values and exact matching for list structures.

🛠️ Project Structure

LongDA/
├── benchmark/ # Benchmark data and documentation
│ ├── benchmark.csv # 505 queries with ground truth
│ └── [SURVEY]/ # Survey-specific folders
│ ├── data/ # Raw data files
│ └── docs/ # Long documentation (codebooks, guides, etc.)
├── configs/ # Model configuration templates
├── tools/ # Custom tools for the agent framework
├── main.py # Main evaluation script
├── evaluate_results.py # Post-hoc evaluation and analysis
├── metric.py # Evaluation metrics implementation
├── my_agent.py # LongTA agent framework
└── utils.py # Utility functions

📚 Citation

If you use LongDA in your research, please cite:

@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}
}

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

📧 Contact

For questions or issues, please:

• Open a GitHub issue
• Contact: yli62@nd.edu

---

Note: This benchmark is for research purposes only. Please comply with data usage policies when using the survey data.