fahmidiqbal/snli_with_instructions

SNLI with Instructions

Dataset Summary

SNLI with Instructions is an instruction-formatted version of the Stanford Natural Language Inference (SNLI) dataset. Each example contains a premise–hypothesis pair embedded in an explicit natural-language inference (NLI) classification prompt, together with its original gold label.

The prompt specifies the direction of inference, the meanings of the three labels, and an exact-output constraint. The purpose is to support instruction-based fine-tuning and controlled studies of robustness and generalization.

This repository contains a processed version, not a new independently annotated dataset. It uses the original SNLI examples and labels, except for invalid-label filtering and the balanced test subset described below.

Dataset Structure

Each JSONL record has three string fields:

Field	Description
id	Unique identifier assigned during preprocessing (for example, train_000001).
text	Full NLI instruction and marked premise–hypothesis pair.
label	Gold label: ENTAILMENT, NEUTRAL, or CONTRADICTION.

The gold label is stored separately in label, not appended to text as the answer.

Label Definitions

• ENTAILMENT: The hypothesis is definitely true given the premise.
• NEUTRAL: The hypothesis might be true or false given the premise.
• CONTRADICTION: The hypothesis is definitely false given the premise.

Prompt Format

Every text field follows this template:

You are a system for natural language inference classification.

Input:
Two marked spans:
- <p>...</p> = PREMISE
- <h>...</h> = HYPOTHESIS

Rules:
- Determine the relationship of the HYPOTHESIS with respect to the PREMISE.
- Use only the information provided in the PREMISE.
- Do NOT assume information that is not stated or logically implied.
- Do NOT reverse the direction of inference.

Labels:
- ENTAILMENT = The hypothesis is definitely true given the premise.
- NEUTRAL = The hypothesis might be true or false given the premise.
- CONTRADICTION = The hypothesis is definitely false given the premise.

Input Pair:
<p> {premise} </p>
<h> {hypothesis} </h>

Output EXACTLY one label: ENTAILMENT, CONTRADICTION, or NEUTRAL.
DO NOT output anything else.

{premise} and {hypothesis} are replaced with the original SNLI sentences. The exact whitespace around those sentences depends on the preprocessing template used to generate the uploaded files.

Dataset Splits

Split	Examples	Processing
Train	All examples with valid labels	Keep original split; convert to instruction format.
Validation	All examples with valid labels	Keep original split; convert to instruction format.
Test	1,500	Sample 500 unique pairs per label without replacement, then shuffle.

The test subset is balanced: 500 ENTAILMENT, 500 NEUTRAL, and 500 CONTRADICTION. Sampling and shuffling use random seed 42. The original full SNLI test split contains 10,000 examples; this repository includes the selected 1,500-example subset instead.

Original Dataset, License, and Citation

• Original dataset: Stanford Natural Language Inference (SNLI)
• Original Hugging Face repository: stanfordnlp/snli
• Paper: A large annotated corpus for learning natural language inference (EMNLP 2015)
• License: Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0)