Pankaj8922/nli-high-quality

NLI High-Quality Balanced Dataset

A combined, filtered, and class-balanced natural language inference (NLI) dataset built from MNLI, SNLI, FEVER-NLI, and ANLI, intended for fine-tuning NLI models for use in zero-shot text classification via the entailment trick (hypothesis = "This example is about {label}.").

The goal of this dataset was quality and generalization over raw volume: rather than concatenating the four source datasets as-is, several filtering stages were applied to remove ambiguous, likely-mislabeled, and shortcut-solvable examples before balancing and splitting.

Dataset structure

Split	Rows
train	638,647
validation	35,480
test	35,481
total	709,608

Columns

Column	Type	Description
premise	string	The premise / context sentence
hypothesis	string	The hypothesis sentence being judged against the premise
label	int (0/1/2)	0 = entailment, 1 = neutral, 2 = contradiction

Classes are balanced: each split has an equal (or near-equal, after stratified splitting) number of examples per label.

How this dataset was built

Source data was combined from four datasets, then passed through the following pipeline, in order:

1. Annotator-agreement filtering (MNLI, SNLI only). Both datasets collect 5 independent annotations per example alongside the majority-vote gold label. Examples where fewer than 4/5 annotators agreed with the gold label were removed, since a low-agreement example usually means the premise/hypothesis pair is genuinely ambiguous or the gold label itself is questionable. FEVER-NLI and ANLI don't have per-annotator labels and passed through this step unfiltered.

2. Deduplication. Combined all four sources, then removed exact-match duplicate (premise, hypothesis) pairs after normalizing whitespace and case.

3. Teacher-model confidence filtering. Every remaining example was scored with MoritzLaurer/DeBERTa-v3-base-mnli-fever-anli, a strong pretrained NLI model. Examples were kept only if the teacher's predicted label matched the gold label with confidence >= 0.9. This removes likely mislabeled or low-signal examples, at the cost of biasing the remaining data toward whatever this specific teacher model finds unambiguous.

4. Hypothesis-only artifact filtering. A simple TF-IDF + logistic regression classifier was trained on the hypothesis sentence alone (the premise was never shown to it), using cross-validated out-of-fold predictions. Examples this shortcut classifier solved correctly with confidence >= 0.7 were removed, since these are solvable via lexical pattern-matching (e.g. negation words correlating with "contradiction") rather than genuine reasoning over the premise. This step specifically targets the annotation artifacts documented in Gururangan et al. 2018 ("Annotation Artifacts in Natural Language Inference Data").

5. Class balancing. Each label was undersampled to match the smallest class count, so the model isn't biased toward predicting whichever label happened to be most common after the filtering steps above.

6. Stratified train/validation/test split (90/5/5), stratified by label.

Known limitations

• The teacher-confidence filtering step (3) biases this dataset toward agreement with one specific model family (DeBERTa-v3 trained on MNLI+FEVER+ANLI). Examples that a genuinely different model or approach might get right, but this teacher finds ambiguous, were removed. This dataset is therefore not fully independent of that teacher's judgments.
• Class balancing was done via undersampling, not synthetic augmentation or loss reweighting, so some correctly labeled majority-class examples were discarded purely to match the smallest class's count.
• No contamination check has been run against MNLI-mismatched, SNLI-test, or ANLI-test splits. If you plan to report zero-shot benchmark numbers on those specific test sets, check for overlap with this training data first.

Intended use

Fine-tuning NLI or zero-shot text classification models. Not intended as a general-purpose sentence-pair dataset outside the entailment/NLI framing.