Need help with anli?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

244 Stars 27 Forks Other 81 Commits 1 Opened issues


Adversarial Natural Language Inference Benchmark

Services available


Need anything else?

Contributors list

Adversarial NLI


Adversarial NLI: A New Benchmark for Natural Language Understanding


Version 1.0 is available here:


The dataset files are all in JSONL format (one JSON per line). Below is one example (in JSON format) with self-explanatory fields.
Note that each example (each line) in the files contains a

field represents a unique id across all the examples in all there rounds of ANLI.
    "uid": "8a91e1a2-9a32-4fd9-b1b6-bd2ee2287c8f", 
    "premise": "Javier Torres (born May 14, 1988 in Artesia, California) is an undefeated Mexican American professional boxer in the Heavyweight division. 
                Torres was the second rated U.S. amateur boxer in the Super Heavyweight division and a member of the Mexican Olympic team.", 
    "hypothesis": "Javier was born in Mexico", 
    "label": "c", 
    "reason": "The paragraph states that Javier was born in the California, US."


AdversarialNLI dataset contains a reason field for each examples in the

split and for some examples in the
split. The reason is collected by asking annotator "Please write a reason for your statement belonging to the category and why you think it was difficult for the system.".


If you want to have your model added to the leaderboard, please reach out to us or submit a PR.


Publication A1 A2 A3
InfoBERT (RoBERTa Large) Wang et al., 2020 75.5 51.4 49.8
ALUM (RoBERTa Large) Liu et al., 2020 72.3 52.1 48.4
GPT-3 Brown et al., 2020 36.8 34.0 40.2
ALBERT ( using the checkpoint in this codebase ) Lan et al., 2019 73.6 58.6 53.4
XLNet Large Yang et al., 2019 67.6 50.7 48.3
RoBERTa Large Liu et al., 2019 73.8 48.9 44.4
BERT Large Devlin et al., 2018 57.4 48.3 43.5

(Updated on Jan 21 2021: The three entries at the bottom show the test set numbers from Table 3 in the ANLI paper. We recommend that you report test set results in your paper. Dev scores, obtained for the models in this code base, are reported below.)


To facilitate research in the field of NLI, we provide an easy-to-use codebase for NLI data preparation and modeling. The code is built upon Transformers with a special focus on NLI.

We welcome researchers from various fields (linguistics, machine learning, cognitive science, psychology, etc.) to try NLI. You can use the code to reproduce the results in our paper or even as a starting point for your research.

Please read more in Start your NLI research.

An important detail in our experiments is that we combine SNLI+MNLI+FEVER-NLI and up-sample different rounds of ANLI to train the models.
We highly recommend you refer to the above link for reproducing the results and training your models such that the results will be comparable to the ones on the leaderboard.

Pre-trained Models

Pre-trained NLI models can be easily called through huggingface model hub.

Version information:

transformers==3.0.2 or later (tested: 3.0.2, 3.1.0, 4.0.0)



The training data is a combination of

ANLI (R1, R2, R3)
. Please also cite the datasets if you are using the pre-trained model.

Please try the code snippet below. ```python from transformers import AutoTokenizer, AutoModelForSequenceClassification import torch

if name == 'main': max_length = 256

premise = "Two women are embracing while holding to go packages."
hypothesis = "The men are fighting outside a deli."

hg_model_hub_name = "ynie/roberta-large-snli_mnli_fever_anli_R1_R2_R3-nli"

hg_model_hub_name = "ynie/albert-xxlarge-v2-snli_mnli_fever_anli_R1_R2_R3-nli"

hg_model_hub_name = "ynie/bart-large-snli_mnli_fever_anli_R1_R2_R3-nli"

hg_model_hub_name = "ynie/electra-large-discriminator-snli_mnli_fever_anli_R1_R2_R3-nli"

hg_model_hub_name = "ynie/xlnet-large-cased-snli_mnli_fever_anli_R1_R2_R3-nli"

tokenizer = AutoTokenizer.from_pretrained(hg_model_hub_name) model = AutoModelForSequenceClassification.from_pretrained(hg_model_hub_name)

tokenized_input_seq_pair = tokenizer.encode_plus(premise, hypothesis, max_length=max_length, return_token_type_ids=True, truncation=True)

input_ids = torch.Tensor(tokenized_input_seq_pair['input_ids']).long().unsqueeze(0)

remember bart doesn't have 'token_type_ids', remove the line below if you are using bart.

token_type_ids = torch.Tensor(tokenized_input_seq_pair['token_type_ids']).long().unsqueeze(0) attention_mask = torch.Tensor(tokenized_input_seq_pair['attention_mask']).long().unsqueeze(0)

outputs = model(input_ids, attention_mask=attention_mask, token_type_ids=token_type_ids, labels=None)


"id2label": {

"0": "entailment",

"1": "neutral",

"2": "contradiction"


predicted_probability = torch.softmax(outputs[0], dim=1)[0].tolist() # batch_size only one

print("Premise:", premise) print("Hypothesis:", hypothesis) print("Entailment:", predicted_probability[0]) print("Neutral:", predicted_probability[1]) print("Contradiction:", predicted_probability[2])

If you are using our pre-trained model checkpoints with the above code snippet, you would expect to get the following numbers.

Huggingface Model Hub Checkpoint A1 (dev) A2 (dev) A3 (dev) A1 (test) A2 (test) A3 (test)
ynie/roberta-large-snli_mnli_fever_anli_R1_R2_R3-nli 73.8 50.8 46.1 73.6 49.3 45.5
ynie/xlnet-large-cased-snli_mnli_fever_anli_R1_R2_R3-nli 73.4 52.3 50.8 70.0 51.4 49.8
ynie/albert-xxlarge-v2-snli_mnli_fever_anli_R1_R2_R3-nli 76.0 57.0 57.0 73.6 58.6 53.4

More in here.


When using this dataset, we ask that you obey some very simple rules:

  1. We want to make it easy for people to provide ablations on test sets without being rate limited, so we release labeled test sets with this distribution. We trust that you will act in good faith, and will not tune on the test set (this should really go without saying)! We may release unlabeled test sets later.

  2. Training data is for training, development data is for development, and test data is for reporting test numbers. This means that you should not e.g. train on the train+dev data from rounds 1 and 2 and then report an increase in performance on the test set of round 3.

  3. We will host a leaderboard on this page. If you want to be added to the leaderboard, please contact us and/or submit a PR with a link to your paper, a link to your code in a public repository (e.g. Github), together with the following information: number of parameters in your model, data used for (pre-)training, and your dev and test results for each round, as well as the total over all rounds.

Other NLI Reference

We used following NLI resources in training the backend model of the adversarial collection:


@inproceedings{nie-etal-2020-adversarial, 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", booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics", year = "2020", publisher = "Association for Computational Linguistics", } ```


ANLI is licensed under Creative Commons-Non Commercial 4.0. See the LICENSE file for details.

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.