fran-gen/snli-smoke-test

SNLI Smoke Test

Dataset Summary

This dataset is a small smoke-test subset derived from the Stanford Natural Language Inference (SNLI) training split. It is intended for fast end-to-end checks of prompt formatting, model adapters, output parsing, and metric pipelines in entailment-lab.

The dataset contains 100 sentence pairs:

• 34 entailment
• 33 contradiction
• 33 neutral

Most of the dataset is organized as complete captionID triplets, where the same premise group contributes one example for each label. This makes the set useful for verifying that a model can distinguish labels without relying on premise changes alone.

Intended Use

Use this dataset for:

• local smoke tests
• CI checks
• parser and metric validation
• quick comparison runs before larger evaluations

Do not use it as a serious benchmark or for model ranking.

Dataset Structure

Each row is stored as JSON Lines with this structure:

{
  "id": "snli_smoke_001",
  "premise": "A person on a horse jumps over a broken down airplane.",
  "hypothesis": "A person is outdoors, on a horse.",
  "label": "entailment",
  "metadata": {
    "source": "snli",
    "source_split": "train",
    "source_pair_id": "3416050480.jpg#4r1e",
    "caption_id": "3416050480.jpg#4",
    "phenomenon": "untyped_smoke_example",
    "reasoning_type": "generic_nli",
    "difficulty": "easy",
    "annotator_labels": ["entailment"],
    "license": "cc-by-sa-4.0",
    "subset": "smoke_test",
    "selection_strategy": "caption_triplet"
  }
}

Source Data

This dataset is derived from SNLI:

• Source dataset: Stanford Natural Language Inference (SNLI) Corpus
• Source split: train
• Original license: CC BY-SA 4.0
• Original dataset page: https://nlp.stanford.edu/projects/snli/

Curation Notes

The subset was deterministically selected to be:

• balanced as closely as possible across the three NLI labels
• small enough for fast local evaluation
• relatively clear and low-ambiguity by selecting rows with a single annotator label matching the gold label
• structured mostly as same-premise triplets across the three labels

The phenomenon, reasoning_type, and selection_strategy fields are lightweight project-level annotations added for smoke-test analysis. They are not part of the original SNLI release.

Limitations

• It is far too small for model comparison.
• It inherits known SNLI limitations and annotation artifacts.
• The semantic tags in metadata are pragmatic project annotations, not gold linguistic labels from SNLI.

Citation

If you use the underlying source data, cite SNLI:

@inproceedings{bowman-etal-2015-large,
  title = "A large annotated corpus for learning natural language inference",
  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",
  year = "2015"
}