---
library_name: zeroshot_classifier
tags:
- transformers
- sentence-transformers
- zeroshot_classifier
license: mit
datasets:
- claritylab/UTCD
language:
- en
pipeline_tag: text-generation
metrics:
- accuracy
---

# Zero-shot Explicit GPT2

This is a modified GPT2 model. 
It was introduced in the Findings of ACL'23 Paper **Label Agnostic Pre-training for Zero-shot Text Classification** by ***Christopher Clarke, Yuzhao Heng, Yiping Kang, Krisztian Flautner, Lingjia Tang and Jason Mars***. 
The code for training and evaluating this model can be found [here](https://github.com/ChrisIsKing/zero-shot-text-classification/tree/master). 

## Model description

This model is intended for zero-shot text classification. 
It was trained under the generative classification framework via explicit training with the aspect-normalized [UTCD](https://huggingface.co/datasets/claritylab/UTCD) dataset. 

- **Finetuned from model:** [`gpt2-medium`](https://huggingface.co/gpt2-medium)


## Usage

Install our [python package](https://pypi.org/project/zeroshot-classifier/): 
```bash
pip install zeroshot-classifier
```

Then, you can use the model like this:

```python
>>> import torch
>>> from zeroshot_classifier.models import ZsGPT2Tokenizer, ZsGPT2LMHeadModel

>>> training_strategy = 'explicit'
>>> model_name = f'claritylab/zero-shot-{training_strategy}-gpt2'
>>> model = ZsGPT2LMHeadModel.from_pretrained(model_name)
>>> tokenizer = ZsGPT2Tokenizer.from_pretrained(model_name, form=training_strategy)

>>> text = "I'd like to have this track onto my Classical Relaxations playlist."
>>> labels = [
>>>     'Add To Playlist', 'Book Restaurant', 'Get Weather', 'Play Music', 'Rate Book', 'Search Creative Work',
>>>     'Search Screening Event'
>>> ]

>>> inputs = tokenizer(dict(text=text, label_options=labels), mode='inference-sample')
>>> inputs = {k: torch.tensor(v).unsqueeze(0) for k, v in inputs.items()}
>>> outputs = model.generate(**inputs, max_length=128)
>>> decoded = tokenizer.batch_decode(outputs, skip_special_tokens=False)[0]
>>> print(decoded)

<|question|>Which of these choices best describes the following document? : " Play Music ", " Add To Playlist ", " Rate Book ", " Get Weather ", " Book Restaurant ", " Search Screening Event ", " Search Creative Work "<|endoftext|><|text|>I'd like to have this track onto my Classical Relaxations playlist.<|endoftext|><|answer|>Play Media<|endoftext|>
```