Note that the model was trained on binary NLI to predict either "entailment" or "not-entailment". This is specifically designed for zero-shot classification, where the difference between "neutral" and "contradiction" is irrelevant.

The base model is xtremedistil-l6-h256-uncased from Microsoft.

Intended uses & limitations

How to use the model

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "MoritzLaurer/xtremedistil-l6-h256-mnli-fever-anli-ling-binary"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

premise = "I first thought that I liked the movie, but upon second thought it was actually disappointing."
hypothesis = "The movie was good."

input = tokenizer(premise, hypothesis, truncation=True, return_tensors="pt")
output = model(input["input_ids"].to(device))  # device = "cuda:0" or "cpu"
prediction = torch.softmax(output["logits"][0], -1).tolist()
label_names = ["entailment", "not_entailment"]
prediction = {name: round(float(pred) * 100, 1) for pred, name in zip(prediction, label_names)}
print(prediction)

Training data

This model was trained on 782 357 hypothesis-premise pairs from 4 NLI datasets: MultiNLI, Fever-NLI, LingNLI and ANLI.

Training procedure

xtremedistil-l6-h256-mnli-fever-anli-ling-binary was trained using the Hugging Face trainer with the following hyperparameters.

training_args = TrainingArguments(
    num_train_epochs=5,              # total number of training epochs
    learning_rate=2e-05,
    per_device_train_batch_size=32,   # batch size per device during training
    per_device_eval_batch_size=32,    # batch size for evaluation
    warmup_ratio=0.1,                # number of warmup steps for learning rate scheduler
    weight_decay=0.06,               # strength of weight decay
    fp16=True                        # mixed precision training
)

Eval results

The model was evaluated using the binary test sets for MultiNLI, ANLI, LingNLI and the binary dev set for Fever-NLI (two classes instead of three). The metric used is accuracy.

dataset	mnli-m-2c	mnli-mm-2c	fever-nli-2c	anli-all-2c	anli-r3-2c	lingnli-2c
accuracy	0.897	0.898	0.861	0.607	0.62	0.827
speed (text/sec, GPU Tesla P100, 128 batch)	1490	1485	760	1186	1062	1791

Limitations and bias

Please consult the original paper and literature on different NLI datasets for potential biases.

BibTeX entry and citation info

If you want to cite this model, please cite the original paper, the respective NLI datasets and include a link to this model on the Hugging Face hub.

Ideas for cooperation or questions?

If you have questions or ideas for cooperation, contact me at m{dot}laurer{at}vu{dot}nl or LinkedIn

Debugging and issues

Note that the model was released recently and older versions of HF Transformers seem to have issues running the model (e.g. resulting in an issue with the tokenizer). Using Transformers==4.13 might solve some issues.