Dataset · allenai
Science test-based textual entailment dataset
allenai/scitail
English premise-hypothesis pairs from science school exams and web texts for NLI.
What's inside
A corpus for checking text logical entailment (entailment) from the AllenAI institute with a volume of 100,000 to 1 million records in English. Hypotheses are formed from multiple-choice science test questions and correct answers. Premises are selected using information retrieval from a large base of web sentences and annotated by crowd workers.
What you can do with it
- Training and evaluating textual entailment (NLI) models in the scientific domain
- Checking factual consistency of language model answers with context
- Fine-tuning science question answering search systems
- Languages
- English
- Size
- 100K–1M rows
- Downloads this month
- 23.8K
- Likes
- 5
Models trained on it
- samratduttaofficial/WaterSheepZero shot classification 87
- xxparthparekhxx/openjev-flashText classification 65
- parsecai/ragzillaText classification 56
- ITLL/WaterSheepZero shot classification 32
- bobox/DeBERTaV3-small-SenTra-AdaptiveLayers-AllSoft-LowTempSentence similarity 22
- bobox/DeBERTa3-base-STr-CosineWavesSentence similarity 20
- anish-shrestha/qwen3.5-0.8b-decision-distill-4bMultiple choice 18
- bobox/DeBERTa-ST-AllLayers-testingSentence similarity 17
How to load it
from datasets import load_dataset
ds = load_dataset("allenai/scitail")