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 DeBERTa-v3-xsmall from Microsoft. The v3 variant of DeBERTa substantially outperforms previous versions of the model by including a different pre-training objective, see the DeBERTa-V3 paper.

For highest performance (but less speed), I recommend using https://huggingface.co/MoritzLaurer/DeBERTa-v3-large-mnli-fever-anli-ling-wanli.

Intended uses & limitations

How to use the model

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")

model_name = "MoritzLaurer/DeBERTa-v3-xsmall-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

DeBERTa-v3-xsmall-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	0.925	0.922	0.892	0.676	0.665	0.888
speed (text/sec, CPU, 128 batch)	6.0	6.3	3.0	5.8	5.0	7.6
speed (text/sec, GPU Tesla P100, 128 batch)	473	487	230	390	340	586

Limitations and bias

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

Citation

If you use this model, please cite: Laurer, Moritz, Wouter van Atteveldt, Andreu Salleras Casas, and Kasper Welbers. 2022. ‘Less Annotating, More Classifying – Addressing the Data Scarcity Issue of Supervised Machine Learning with Deep Transfer Learning and BERT - NLI’. Preprint, June. Open Science Framework. https://osf.io/74b8k.

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 DeBERTa-v3 was released on 06.12.21 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.