dleemiller/BigCat-NLI

Big Concatenation (BigCat) NLI

Update 10-19-2025

Removed FOL-NLI after initial training test showed poor compatibility with the other datasets. It would likely be better to train it separately if that task is desired.

Overview

This dataset is a comprehensive compilation of 6 major Natural Language Inference (NLI) datasets, unified into a single standardized format. All datasets have been aligned to use consistent 3-label classification (entailment, neutral, contradiction), deduplicated, and checked for contamination/leakage.

Total Examples: ~2.43M

• Training Set: ~2.39M examples
• Validation Set: 10k examples
• Test Set: 35.6k examples

Label Format

All examples use integer labels following this standard mapping:

• 0 = Entailment - The hypothesis logically follows from the premise
• 1 = Neutral - The hypothesis may or may not follow from the premise
• 2 = Contradiction - The hypothesis contradicts the premise

Dataset Schema

Each example contains:

• premise (string): The premise text
• hypothesis (string): The hypothesis text
• label (int64): Classification label (0, 1, or 2)
• source (string): Original source dataset identifier

Source Datasets

This compilation includes data from the following sources:

1. NLI-FEVER

Source: pietrolesci/nli_fever

A Natural Language Inference version of the FEVER (Fact Extraction and VERification) dataset. Columns premise and hypothesis have been swapped from the original dataset to better directionally align with traditional NLI models.

Citation:

@inproceedings{Thorne18Fever,
  author = {Thorne, James and Vlachos, Andreas and Christodoulopoulos, Christos and Mittal, Arpit},
  title = {{FEVER}: a Large-scale Dataset for Fact Extraction and {VERification}},
  booktitle = {NAACL-HLT},
  year = {2018}
}

2. ANLI (Adversarial NLI)

Source: facebook/anli

Includes all three rounds (R1, R2, R3) of the Adversarial Natural Language Inference dataset.

Citation:

@inproceedings{Nie2020AdversarialNA,
  title={Adversarial NLI: A New Benchmark for Natural Language Understanding},
  author={Yixin Nie and Adina Williams and Emily Dinan and Mohit Bansal and Jason Weston and Douwe Kiela},
  booktitle={ACL},
  year={2020}
}

3. SNLI (Stanford Natural Language Inference)

Source: Generating Data to Mitigate Spurious Correlations in Natural Language Inference Datasets (train split)

Source: stanfordnlp/snli (val/test splits)

The Stanford Natural Language Inference corpus.

Citation:

@inproceedings{snli:emnlp2015,
  author = {Bowman, Samuel R. and Angeli, Gabor and Potts, Christopher and Manning, Christopher D.},
  booktitle = {Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  publisher = {Association for Computational Linguistics},
  title = {A large annotated corpus for learning natural language inference},
  year = {2015}
}

4. MNLI (Multi Natural Language Inference)

Source: Generating Data to Mitigate Spurious Correlations in Natural Language Inference Datasets (train split)

Source: nyu-mll/glue (mnli subset, val/test splits)

MNLI from the GLUE benchmark.

Citation:

@inproceedings{N18-1101,
  author = {Williams, Adina and Nangia, Nikita and Bowman, Samuel},
  booktitle = {Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers)},
  title = {A Broad-Coverage Challenge Corpus for Sentence Understanding through Inference},
  year = {2018}
}

5. LingNLI

Source: tasksource/lingnli

Linguistically-informed Natural Language Inference dataset.

Citation:

@misc{TaskSource,
  author = {Jiang, Jimin},
  title = {TaskSource NLI Collections},
  year = {2023},
  howpublished = {\url{https://huggingface.co/tasksource}}
}

6. WANLI (Worker-AI Collaboration for NLI)

Source: alisawuffles/WANLI

A dataset created through worker and AI collaboration.

Citation:

@inproceedings{liu-etal-2022-wanli,
  title = "{WANLI}: Worker and AI Collaboration for Natural Language Inference Dataset Creation",
  author = "Liu, Alisa and Swayamdipta, Swabha and Smith, Noah A. and Choi, Yejin",
  booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2022",
  year = "2022"
}

Data Processing

Label Alignment

Each source dataset was carefully inspected and mapped to the standard label format:

• NLI-FEVER was remapped from SUPPORTS/REFUTES/NOT ENOUGH INFO to entailment/contradiction/neutral
• String labels were converted to integer format where applicable
• Examples with invalid labels (e.g., -1) were filtered out

Split Creation

After deduplication, new train/validation/test splits were created:

1. All data from all source datasets was concatenated (separately by split)
2. The combined train dataset was shuffled
3. Full test sets concatenated
4. 12,000 examples were randomly sampled for the validation set
5. Remove concatenated premise/hypothesis rows with token_count >460 (length consistency)

License

This compiled dataset is subject to the licenses of all constituent datasets.

Users should review and comply with the individual licenses:

• SNLI: CC BY-SA 4.0
• MNLI: CC BY-SA 4.0
• ANLI: CC BY-NC 4.0
• Other datasets: Refer to individual dataset pages

Acknowledgments

We thank the creators and contributors of all source datasets for making their work available. This compilation would not be possible without their efforts to advance Natural Language Understanding research.

Citation

If you use this compiled dataset, please cite all relevant source datasets as listed above, as well as this compilation:

@misc{nli-compiled-2025,
  title = {BigCat NLI Dataset},
  author = {Lee Miller},
  year = {2025},
  howpublished = {Compilation of 6 major NLI datasets}
}