CausalBN-Bench Dataset

This dataset package contains the benchmark assets for CausalBN-Bench: A Comprehensive Benchmark for Causal Learning Capability of LLMs.

CausalBN-Bench evaluates large language models on Bayesian-network-based causal learning tasks:

• Correlation identification
• Causal skeleton identification
• Causality identification

Contents

• data/source_networks_and_labels/: raw Bayesian network files and ground-truth labels.
• data/main_tasks/correlation_identification/: prompt files for the paper correlation identification task.
• data/main_tasks/causal_skeleton_identification/: prompt files, labels, and network files for the paper causal skeleton identification task.
• data/main_tasks/causality_identification/: prompt files for the paper causality identification task, including direct-causality and variable-name-only variants.
• data/prompt_formats/background_knowledge/: background-knowledge files and combined background-knowledge prompts.
• data/appendix/variable_refactorization/: variable-modification/refactorization prompts and labels used for robustness or appendix analyses.
• data/appendix/causal_strength/ranking/: causal-strength ranking prompts for the appendix/future-direction exploration.
• metadata/file_manifest.csv: released data-file manifest with sizes.
• metadata/checksums.sha256: SHA256 checksum file for released data files.
• metadata/excluded_pickles_manifest.csv: pickle files intentionally excluded from the public dataset package.

Paper Task Mapping

Paper benchmark component	Released dataset location	Notes
Correlation identification	data/main_tasks/correlation_identification/	Pairwise relatedness prompts.
Causal skeleton identification	data/main_tasks/causal_skeleton_identification/	Conditional-dependence prompts, labels, and BIF/network files.
Causality identification	data/main_tasks/causality_identification/direct_causality_prompts/ and data/main_tasks/causality_identification/variable_name_prompts/	Direct-cause and variable-name prompt variants.
Prompt format with background knowledge	data/prompt_formats/background_knowledge/	Knowledge files and combined background-knowledge prompts.
Variable refactorization / modified variable names	data/appendix/variable_refactorization/questions/, data/appendix/variable_refactorization/labels/, data/appendix/variable_refactorization/nested/	Appendix robustness/ablation materials.
Causal strength exploration	data/appendix/causal_strength/ranking/	Appendix/future-direction ranking prompts, not one of the three main benchmark tasks.
Raw networks and labels	data/source_networks_and_labels/	Source Bayesian-network structures and generated labels.

Dataset Size Summary

This staged release contains 583 released data files, plus metadata files. The largest sections are full question/prompt files:

• data/main_tasks/causality_identification/direct_causality_prompts/: about 6.57 GiB
• data/main_tasks/causality_identification/variable_name_prompts/: about 5.02 GiB
• data/main_tasks/correlation_identification/: about 5.02 GiB
• data/appendix/variable_refactorization/questions/: about 5.15 GiB
• data/appendix/causal_strength/ranking/: about 1.31 GiB

See metadata/summary_by_section.csv and metadata/file_counts_by_extension.csv for exact file counts and byte sizes.

Publishing To Hugging Face

Create a dataset repository on Hugging Face, then run:

pip install -U huggingface_hub
huggingface-cli login
python scripts/upload_to_hf.py your-org/CausalBN-Bench
# or: bash scripts/upload_to_hf.sh your-org/CausalBN-Bench

Large text/CSV/network files are marked for Git LFS in .gitattributes.

Notes On Pickle Files

Large *_full.pkl files from the working directory were intentionally excluded. They are very large, Python-specific, and less suitable for public dataset release. The corresponding CSV/TXT files are included where available. See metadata/excluded_pickles_manifest.csv for the excluded file list.

Usage

After downloading the dataset, the CSV/TXT files can be used directly by the source code repository. For example, causality-identification prompts are under:

data/main_tasks/causality_identification/direct_causality_prompts/

The GitHub source-code release should link to this Hugging Face dataset and use scripts/download_data.py to fetch it.

Citation

@ARTICLE{causalbnbench,
  author={Zhou, Yu and Wu, Xingyu and Wu, Jibin and Feng, Liang and Tan, Kay Chen},
  journal={IEEE Transactions on Artificial Intelligence},
  title={CausalBN-Bench: A Comprehensive Benchmark for Causal Learning Capability of LLMs},
  year={2026},
  volume={},
  number={},
  pages={1-15},
  doi={10.1109/TAI.2026.3703427}
}