language:
- en
tags:
- text-classification
- zero-shot-classification
license: mit
metrics:
- accuracy
datasets:
- multi_nli
- anli
- fever
- lingnli
- alisawuffles/WANLI
widget:
- text: >-
I first thought that I really liked the movie, but upon second thought it
was actually disappointing. [SEP] The movie was good.
model-index:
- name: DeBERTa-v3-large-mnli-fever-anli-ling-wanli
results:
- task:
type: text-classification
name: Natural Language Inference
dataset:
type: multi_nli
name: MultiNLI-matched
split: validation_matched
metrics:
- type: accuracy
value: 0,912
verified: false
- task:
type: text-classification
name: Natural Language Inference
dataset:
type: multi_nli
name: MultiNLI-mismatched
split: validation_mismatched
metrics:
- type: accuracy
value: 0,908
verified: false
- task:
type: text-classification
name: Natural Language Inference
dataset:
type: anli
name: ANLI-all
split: test_r1+test_r2+test_r3
metrics:
- type: accuracy
value: 0,702
verified: false
- task:
type: text-classification
name: Natural Language Inference
dataset:
type: anli
name: ANLI-r3
split: test_r3
metrics:
- type: accuracy
value: 0,64
verified: false
- task:
type: text-classification
name: Natural Language Inference
dataset:
type: alisawuffles/WANLI
name: WANLI
split: test
metrics:
- type: accuracy
value: 0,77
verified: false
- task:
type: text-classification
name: Natural Language Inference
dataset:
type: lingnli
name: LingNLI
split: test
metrics:
- type: accuracy
value: 0,87
verified: false
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 NLI and zero-shot model on the Hugging Face Hub as of 06.06.22. It significantly outperforms all other large models on the ANLI benchmark.
The foundation model is DeBERTa-v3-large from Microsoft. Released on 06.12.21, DeBERTa-v3-large is currently the best large-sized foundation model for text classification. It combines several recent innovations compared to classical Masked Language Models like BERT, RoBERTa etc., see the paper
Intended uses & limitations
How to use the model
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
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 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.
BibTeX entry and citation info
If you want to cite this model, please cite my preprint on low-resource text classification and the original DeBERTa-v3 paper.
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.