Lidor-Mashiach/mnli-training-paraphrase-augmentation

MNLI Training Paraphrase Augmentation

Purpose

This dataset contains new paraphrases created for training paraphrase augmentation during Phase B of the research project.

It was not used as an NLI evaluation set.

It was not used for paraphrase consistency evaluation.

It is separate from the published MNLI Paraphrase Bank used for evaluation. The generation records report zero collisions with that evaluation bank.

The CSV contains only the new augmentation rows. It does not contain the full original MNLI training split. To create an augmented training set, combine these rows with the original MNLI training data.

Files

File	Purpose
mnli-training-paraphrase-augmentation.csv	New paraphrases created for Phase B training
mnli-training-paraphrase-augmentation-stats.json	Merged generation statistics and experimental budgets
README.md	Dataset documentation

Columns

Column	Meaning
pair_id	Identifier of the original MNLI training pair
premise	Original premise
hypothesis	Original hypothesis
paraphrase	New hypothesis paraphrase created for augmentation
label	Original NLI label
para_idx	Index of the new paraphrase for the same hypothesis

How the data was created

The augmentation rows were generated in three parallel parts and then merged into one CSV.

Candidate paraphrases were generated with humarin/chatgpt_paraphraser_on_T5_base.

Candidates were checked with MoritzLaurer/DeBERTa-v3-large-mnli-fever-anli-ling-wanli.

A candidate was kept only when it passed the following checks:

1. Its length stayed within the allowed ratio relative to the original hypothesis.
2. The candidate and original hypothesis were predicted to entail one another.
3. The NLI relation between the premise and candidate matched the original gold label.
4. The candidate was not a duplicate.
5. The candidate did not appear in the published evaluation paraphrase bank.

The target was up to five new paraphrases per requested hypothesis.

Statistics

Measure	Value
Hypotheses requested	319,317
Hypotheses with at least one new paraphrase	318,542
Hypotheses with no new paraphrase	775
Hypotheses reaching five new paraphrases	310,128
New paraphrase rows	1,573,668
Parallel generation parts	3
Collisions with the evaluation paraphrase bank	0

The merged rejection counts, duplicate checks, and experimental budgets appear in mnli-training-paraphrase-augmentation-stats.json.

Phase B context

The augmentation pool supported three experimental arms:

• targeted
• random
• oracle

The arm budgets were defined separately for BERT Base MNLI and DeBERTa Large MNLI.

These arm names describe the research design. They are not Hugging Face dataset splits.

Intended use

This dataset is intended for model training experiments that add new paraphrase rows to the original MNLI training data.

The original premise, hypothesis, and label are preserved. The new paraphrase is used as the augmented hypothesis.

Researchers should report which arm was used, the number of augmentation rows added, the model configuration, the random seed, and the exact repository version.

Not intended for evaluation

Do not treat this dataset as an independent MNLI validation or test set.

Do not use it as the paraphrase consistency benchmark published in the separate MNLI Paraphrase Bank repository.

Training and evaluation data should remain separated according to the experimental protocol.

Limitations

The paraphrases were generated and checked by models. They were not reviewed one by one by human annotators.

Automated checks can accept an imperfect paraphrase or reject a valid one.

The rows depend on the generator, verifier, filtering rules, and Phase B sampling design.

The CSV contains augmentation rows only. Users must obtain the original MNLI training data separately.

License

MultiNLI contains material from several sources. Its dataset card lists CC BY 3.0, CC BY-SA 3.0, MIT, and other source terms.

This repository uses the same mixed source terms notice.

Users must review the original MultiNLI documentation and follow the terms that apply to the relevant source material.

Associated research

This dataset was created during Phase B of the following research manuscript:

“Opening the Black Box: Localizing semantic inconsistency in NLI models with Deep k -Nearest Neighbors”

The manuscript is in preparation. It has not been submitted or published.

This section will be updated when a public preprint or accepted version becomes available.

Citation

Until the paper is public, please cite this dataset repository:

@misc{mashiach2026mnlitrainingparaphraseaugmentation,
  author = {Lidor Mashiach},
  title = {MNLI Training Paraphrase Augmentation},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/Lidor-Mashiach/mnli-training-paraphrase-augmentation}
}

Please also cite the original MultiNLI paper.

Contact

Questions, corrections, and reproducibility reports can be posted in the Community tab of this repository.