Skip to content

Dataset · facebook

Complex adversarial natural language benchmark

facebook/anli

An English dataset for natural language inference (NLI) built with human and model participation.

What's inside

A large-scale English dataset (100K to 1M examples) for the natural language inference (NLI) task. Collected via an iterative adversarial method involving humans and neural networks across three rounds, each featuring train, dev, and test splits. Designed by the authors to be significantly harder than the classic SNLI and MNLI benchmarks.

What you can do with it

  • Evaluating language model robustness against complex logical traps
  • Fine-tuning classifiers for fact-checking and text contradiction tasks
  • Comparing LLMs on a difficult NLI benchmark from the Facebook team
Tasks
Text classification
Languages
English
Size
100K–1M rows
Downloads this month
31.1K
Likes
59

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("facebook/anli")