DeBERTa-v3-large-mnli-fever-anli-ling-wanli

Model description

This model was fine-tuned on the MultiNLI, Fever-NLI, Adversarial-NLI (ANLI), LingNLI and WANLI datasets, which comprise 885 242 NLI hypothesis-premise pairs. This model is the best performing NLI model on the Hugging Face Hub as of 06.06.22 and can be used for zero-shot classification. It significantly outperforms all other large models on the ANLI benchmark.

The foundation model is DeBERTa-v3-large from Microsoft. DeBERTa-v3 combines several recent innovations compared to classical Masked Language Models like BERT, RoBERTa etc., see the paper

How to use the model

Simple zero-shot classification pipeline

from transformers import pipeline
classifier = pipeline("zero-shot-classification", model="MoritzLaurer/DeBERTa-v3-large-mnli-fever-anli-ling-wanli")
sequence_to_classify = "Angela Merkel is a politician in Germany and leader of the CDU"
candidate_labels = ["politics", "economy", "entertainment", "environment"]
output = classifier(sequence_to_classify, candidate_labels, multi_label=False)
print(output)

NLI use-case

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-large-mnli-fever-anli-ling-wanli"
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 not 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", "neutral", "contradiction"]
prediction = {name: round(float(pred) * 100, 1) for pred, name in zip(prediction, label_names)}
print(prediction)

Training data

DeBERTa-v3-large-mnli-fever-anli-ling-wanli was trained on the MultiNLI, Fever-NLI, Adversarial-NLI (ANLI), LingNLI and WANLI datasets, which comprise 885 242 NLI hypothesis-premise pairs. Note that SNLI was explicitly excluded due to quality issues with the dataset. More data does not necessarily make for better NLI models.

Training procedure

DeBERTa-v3-large-mnli-fever-anli-ling-wanli was trained using the Hugging Face trainer with the following hyperparameters. Note that longer training with more epochs hurt performance in my tests (overfitting).

training_args = TrainingArguments(
    num_train_epochs=4,              # total number of training epochs
    learning_rate=5e-06,
    per_device_train_batch_size=16,   # batch size per device during training
    gradient_accumulation_steps=2,    # doubles the effective batch_size to 32, while decreasing memory requirements
    per_device_eval_batch_size=64,    # batch size for evaluation
    warmup_ratio=0.06,                # number of warmup steps for learning rate scheduler
    weight_decay=0.01,               # strength of weight decay
    fp16=True                        # mixed precision training
)

Eval results

The model was evaluated using the test sets for MultiNLI, ANLI, LingNLI, WANLI and the dev set for Fever-NLI. The metric used is accuracy. The model achieves state-of-the-art performance on each dataset. Surprisingly, it outperforms the previous state-of-the-art on ANLI (ALBERT-XXL) by 8,3%. I assume that this is because ANLI was created to fool masked language models like RoBERTa (or ALBERT), while DeBERTa-v3 uses a better pre-training objective (RTD), disentangled attention and I fine-tuned it on higher quality NLI data.

Datasets	mnli_test_m	mnli_test_mm	anli_test	anli_test_r3	ling_test	wanli_test
Accuracy	0.912	0.908	0.702	0.64	0.87	0.77
Speed (text/sec, A100 GPU)	696.0	697.0	488.0	425.0	828.0	980.0

Limitations and bias

Please consult the original DeBERTa-v3 paper and literature on different NLI datasets for more information on the training data and potential biases. The model will reproduce statistical patterns in the training data.

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.