Model description: deberta-v3-large-zeroshot-v1.1-all-33

The model is designed for zero-shot classification with the Hugging Face pipeline.

The model can do one universal classification task: determine whether a hypothesis is "true" or "not true" given a text (entailment vs. not_entailment).
This task format is based on the Natural Language Inference task (NLI). The task is so universal that any classification task can be reformulated into this task.

A detailed description of how the model was trained and how it can be used is available in this paper.

Training data

The model was trained on a mixture of 33 datasets and 387 classes that have been reformatted into this universal format.

1. Five NLI datasets with ~885k texts: "mnli", "anli", "fever", "wanli", "ling"
2. 28 classification tasks reformatted into the universal NLI format. ~51k cleaned texts were used to avoid overfitting: 'amazonpolarity', 'imdb', 'appreviews', 'yelpreviews', 'rottentomatoes', 'emotiondair', 'emocontext', 'empathetic', 'financialphrasebank', 'banking77', 'massive', 'wikitoxic_toxicaggregated', 'wikitoxic_obscene', 'wikitoxic_threat', 'wikitoxic_insult', 'wikitoxic_identityhate', 'hateoffensive', 'hatexplain', 'biasframes_offensive', 'biasframes_sex', 'biasframes_intent', 'agnews', 'yahootopics', 'trueteacher', 'spam', 'wellformedquery', 'manifesto', 'capsotu'.

See details on each dataset here: https://github.com/MoritzLaurer/zeroshot-classifier/blob/main/datasets_overview.csv

Note that compared to other NLI models, this model predicts two classes (entailment vs. not_entailment) as opposed to three classes (entailment/neutral/contradiction)

The model was only trained on English data. For multilingual use-cases, I recommend machine translating texts to English with libraries like EasyNMT. English-only models tend to perform better than multilingual models and validation with English data can be easier if you don't speak all languages in your corpus.

How to use the model

Simple zero-shot classification pipeline

#!pip install transformers[sentencepiece]
from transformers import pipeline
text = "Angela Merkel is a politician in Germany and leader of the CDU"
hypothesis_template = "This example is about {}"
classes_verbalized = ["politics", "economy", "entertainment", "environment"]
zeroshot_classifier = pipeline("zero-shot-classification", model="MoritzLaurer/deberta-v3-large-zeroshot-v1.1-all-33")
output = zeroshot_classifier(text, classes_verbalised, hypothesis_template=hypothesis_template, multi_label=False)
print(output)

Details on data and training

The code for preparing the data and training & evaluating the model is fully open-source here: https://github.com/MoritzLaurer/zeroshot-classifier/tree/main

Hyperparameters and other details are available in this Weights & Biases repo: https://wandb.ai/moritzlaurer/deberta-v3-large-zeroshot-v1-1-all-33/table?workspace=user-

Metrics

Balanced accuracy is reported for all datasets. deberta-v3-large-zeroshot-v1.1-all-33 was trained on all datasets, with only maximum 500 texts per class to avoid overfitting. The metrics on these datasets are therefore not strictly zeroshot, as the model has seen some data for each task during training. deberta-v3-large-zeroshot-v1.1-heldout indicates zeroshot performance on the respective dataset. To calculate these zeroshot metrics, the pipeline was run 28 times, each time with one dataset held out from training to simulate a zeroshot setup.

Limitations and bias

The model can only do text classification tasks.

Please consult the original DeBERTa paper and the papers for the different datasets for potential biases.

License

The base model (DeBERTa-v3) is published under the MIT license. The datasets the model was fine-tuned on are published under a diverse set of licenses. The following table provides an overview of the non-NLI datasets used for fine-tuning, information on licenses, the underlying papers etc.: https://github.com/MoritzLaurer/zeroshot-classifier/blob/main/datasets_overview.csv

Citation

If you use this model academically, please cite:

@misc{laurer_building_2023,
	title = {Building {Efficient} {Universal} {Classifiers} with {Natural} {Language} {Inference}},
	url = {http://arxiv.org/abs/2312.17543},
	doi = {10.48550/arXiv.2312.17543},
	abstract = {Generative Large Language Models (LLMs) have become the mainstream choice for fewshot and zeroshot learning thanks to the universality of text generation. Many users, however, do not need the broad capabilities of generative LLMs when they only want to automate a classification task. Smaller BERT-like models can also learn universal tasks, which allow them to do any text classification task without requiring fine-tuning (zeroshot classification) or to learn new tasks with only a few examples (fewshot), while being significantly more efficient than generative LLMs. This paper (1) explains how Natural Language Inference (NLI) can be used as a universal classification task that follows similar principles as instruction fine-tuning of generative LLMs, (2) provides a step-by-step guide with reusable Jupyter notebooks for building a universal classifier, and (3) shares the resulting universal classifier that is trained on 33 datasets with 389 diverse classes. Parts of the code we share has been used to train our older zeroshot classifiers that have been downloaded more than 55 million times via the Hugging Face Hub as of December 2023. Our new classifier improves zeroshot performance by 9.4\%.},
	urldate = {2024-01-05},
	publisher = {arXiv},
	author = {Laurer, Moritz and van Atteveldt, Wouter and Casas, Andreu and Welbers, Kasper},
	month = dec,
	year = {2023},
	note = {arXiv:2312.17543 [cs]},
	keywords = {Computer Science - Artificial Intelligence, Computer Science - Computation and Language},
}

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 can have issues running the model (e.g. resulting in an issue with the tokenizer). Using Transformers>=4.13 might solve some issues.

Hypotheses used for classification

The hypotheses in the tables below were used to fine-tune the model. Inspecting them can help users get a feeling for which type of hypotheses and tasks the model was trained on. You can formulate your own hypotheses by changing the hypothesis_template of the zeroshot pipeline. For example:

from transformers import pipeline
text = "Angela Merkel is a politician in Germany and leader of the CDU"
hypothesis_template = "Merkel is the leader of the party: {}"
classes_verbalized = ["CDU", "SPD", "Greens"]
zeroshot_classifier = pipeline("zero-shot-classification", model="MoritzLaurer/deberta-v3-large-zeroshot-v1.1-all-33")
output = zeroshot_classifier(text, classes_verbalized, hypothesis_template=hypothesis_template, multi_label=False)
print(output)

Note that a few rows in the massive and banking77 datasets contain nan because some classes were so ambiguous/unclear that I excluded them from the data.

wellformedquery

biasframes_sex

biasframes_intent

biasframes_offensive

financialphrasebank

rottentomatoes

amazonpolarity

imdb

appreviews

yelpreviews

wikitoxic_toxicaggregated

wikitoxic_obscene

wikitoxic_threat

wikitoxic_insult

wikitoxic_identityhate

hateoffensive

hatexplain

spam

emotiondair

emocontext

empathetic

agnews

yahootopics

massive

banking77

trueteacher

capsotu

manifesto