metadata

language:
  - en
tags:
  - text-classification
  - zero-shot-classification
pipeline_tag: zero-shot-classification
library_name: transformers
license: mit
datasets:
  - nyu-mll/multi_nli
  - fever

Model description: deberta-v3-base-zeroshot-v2.0

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

The main advantage of this zeroshot-v2.0 series of models is that they are trained on commercially-friendly data and are fully commercially usable, while my older zeroshot-v1.1 models included training data with non-commercially licenses data. An overview of the latest zeroshot classifiers with different sizes and licenses is available in my Zeroshot Classifier Collection.

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

Training data

The model is trained on two types of fully commercially-friendly data:

Synthetic data generated with Mixtral-8x7B-Instruct-v0.1. I first created a list of 500+ diverse text classification tasks for 25 professions in conversations with Mistral-large. The data was manually curated. I then used this as seed data to generate several hundred thousand texts for the different tasks with Mixtral-8x7B-Instruct-v0.1. The final dataset used is available in the synthetic_zeroshot_mixtral_v0.1 dataset in the subset mixtral_written_text_for_tasks_v4. Data curation was done in multiple iterations and I will release more information on this process soon.
Two commercially-friendly NLI datasets: (MNLI, FEVER-NLI). These datasets were added to increase generalization. Datasets like ANLI were excluded due to their non-commercial license.

The model was only trained on English data. I will release a multilingual version of this model soon. For multilingual use-cases, I alternatively 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

#!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-base-zeroshot-v2.0")
output = zeroshot_classifier(text, classes_verbalised, hypothesis_template=hypothesis_template, multi_label=False)
print(output)

multi_label=False forces the model to decide on only one class. multi_label=True enables the model to choose multiple classes.

Metrics

The model was evaluated on 28 different text classification tasks with the balanced_accuracy metric. The main reference point is facebook/bart-large-mnli which is at the time of writing (27.03.24) the most used commercially-friendly 0-shot classifier. The different zeroshot-v2.0 models were all trained with the same data and the only difference is the underlying foundation model.

	facebook/bart-large-mnli	roberta-base-zeroshot-v2.0	roberta-large-zeroshot-v2.0	deberta-v3-base-zeroshot-v2.0	deberta-v3-large-zeroshot-v2.0
all datasets mean	0.566	0.612	0.65	0.647	0.697
amazonpolarity (2)	0.937	0.924	0.951	0.937	0.952
imdb (2)	0.892	0.871	0.904	0.893	0.923
appreviews (2)	0.934	0.913	0.937	0.938	0.943
yelpreviews (2)	0.948	0.953	0.977	0.979	0.988
rottentomatoes (2)	0.831	0.803	0.841	0.841	0.87
emotiondair (6)	0.495	0.523	0.514	0.487	0.495
emocontext (4)	0.605	0.535	0.609	0.566	0.687
empathetic (32)	0.366	0.386	0.417	0.388	0.455
financialphrasebank (3)	0.673	0.521	0.445	0.678	0.656
banking77 (72)	0.327	0.138	0.297	0.433	0.542
massive (59)	0.454	0.481	0.599	0.533	0.599
wikitoxic_toxicaggreg (2)	0.609	0.752	0.768	0.752	0.751
wikitoxic_obscene (2)	0.728	0.818	0.854	0.853	0.884
wikitoxic_threat (2)	0.531	0.796	0.874	0.861	0.876
wikitoxic_insult (2)	0.514	0.738	0.802	0.768	0.778
wikitoxic_identityhate (2)	0.567	0.776	0.801	0.774	0.801
hateoffensive (3)	0.41	0.497	0.484	0.539	0.634
hatexplain (3)	0.373	0.423	0.385	0.441	0.446
biasframes_offensive (2)	0.499	0.571	0.587	0.546	0.648
biasframes_sex (2)	0.503	0.703	0.845	0.794	0.877
biasframes_intent (2)	0.635	0.541	0.635	0.562	0.696
agnews (4)	0.722	0.765	0.764	0.694	0.824
yahootopics (10)	0.303	0.55	0.621	0.575	0.605
trueteacher (2)	0.492	0.488	0.501	0.505	0.515
spam (2)	0.523	0.537	0.528	0.531	0.698
wellformedquery (2)	0.528	0.5	0.5	0.5	0.476
manifesto (56)	0.088	0.111	0.206	0.198	0.277
capsotu (21)	0.375	0.525	0.558	0.543	0.631

When to use which model

deberta-v3-zeroshot vs. roberta-zeroshot: deberta-v3 performs clearly better than roberta, but it is slower. roberta is directly compatible with Hugging Face's production inference TEI containers and flash attention. These containers are a good choice for production use-cases. tl;dr: For accuracy, use a deberta-v3 model. If production inference speed is a concern, you can consider a roberta model (e.g. in a TEI container and HF Inference Endpoints).
zeroshot-v1.1 vs. zeroshot-v2.0 models: My zeroshot-v1.1 models (see Zeroshot Classifier Collection)) perform better on these 28 datasets, but they are trained on several datasets with non-commercial licenses. For commercial users, I therefore recommend using a v2.0 model and non-commercial users might get better performance with a v1.1 model.
The latest updates on new models are always available in the Zeroshot Classifier Collection.

Reproduction

Reproduction code is available in the v2_synthetic_data directory here: https://github.com/MoritzLaurer/zeroshot-classifier/tree/main

Limitations and bias

The model can only do text classification tasks.

Biases can come from the underlying foundation model, the human NLI training data and the synthetic data generated by Mixtral.

License

The foundation model (DeBERTa-v3) is published under the MIT license. The training data is released under different, permissive, commercially-friendly licenses (MNLI, FEVER-NLI, synthetic_zeroshot_mixtral_v0.1)

Citation

This model is an extension of the research described in this paper.

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 moritz{at}huggingface{dot}co or LinkedIn

Flexible usage and "prompting"

You can formulate your own hypotheses by changing the hypothesis_template of the zeroshot pipeline. Similar to "prompt engineering" for LLMs, you can test different formulations of your hypothesis_template and verbalized classes to improve performance.

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-base-zeroshot-v2.0")
output = zeroshot_classifier(text, classes_verbalized, hypothesis_template=hypothesis_template, multi_label=False)
print(output)