NLI — one task, three production uses
premise (t):
"A cat is sleeping on the couch."
hypothesis (h):
"There is a cat in the room."
transformer encoder
[CLS] t [SEP] h [SEP]
DeBERTa-v3 / RoBERTa / BART
entailment
0.97
neutral
0.02
contradiction
0.01
same model, three plugs
1. zero-shot classification
t = document
h = "about {label}"
pick max-entailment label
no training data needed
2. faithfulness (RAG)
t = retrieved context
h = generated answer claim
entailment → supported
RAGAS faithfulness core
3. summarization fact-check
t = source document
h = summary sentence
contradiction → hallucination
neutral → unsupported claim