• Description:

Adversarial NLI (ANLI) is a large-scale NLI benchmark dataset, collected via an iterative, adversarial human-and-model-in-the-loop procedure.

    'context': Text(shape=(), dtype=string),
    'hypothesis': Text(shape=(), dtype=string),
    'label': ClassLabel(shape=(), dtype=int64, num_classes=3),
    'uid': Text(shape=(), dtype=string),
  • Feature documentation:
Feature Class Shape Dtype Description
context Text string
hypothesis Text string
label ClassLabel int64
uid Text string
    title = "Adversarial NLI: A New Benchmark for Natural Language Understanding",
    author = "Nie, Yixin and
      Williams, Adina and
      Dinan, Emily  and
      Bansal, Mohit and
      Weston, Jason and
      Kiela, Douwe",
    url =""

anli/r1 (default config)

  • Config description: Round One

  • Dataset size: 9.04 MiB

  • Splits:

Split Examples
'test' 1,000
'train' 16,946
'validation' 1,000


  • Config description: Round Two

  • Dataset size: 22.39 MiB

  • Splits:

Split Examples
'test' 1,000
'train' 45,460
'validation' 1,000


  • Config description: Round Three

  • Dataset size: 47.03 MiB

  • Splits:

Split Examples
'test' 1,200
'train' 100,459
'validation' 1,200