Zero-shot SELECTRA: A zero-shot classifier based on SELECTRA
Zero-shot SELECTRA is a SELECTRA model fine-tuned on the Spanish portion of the XNLI dataset. You can use it with Hugging Face's Zero-shot pipeline to make zero-shot classifications.
In comparison to our previous zero-shot classifier based on BETO, zero-shot SELECTRA is much more lightweight. As shown in the Metrics section, the small version (5 times fewer parameters) performs slightly worse, while the medium version (3 times fewer parameters) outperforms the BETO based zero-shot classifier.
Usage
from transformers import pipeline
classifier = pipeline("zero-shot-classification",
model="Recognai/zeroshot_selectra_medium")
classifier(
"El autor se perfila, a los 50 años de su muerte, como uno de los grandes de su siglo",
candidate_labels=["cultura", "sociedad", "economia", "salud", "deportes"],
hypothesis_template="Este ejemplo es {}."
)
"""Output
{'sequence': 'El autor se perfila, a los 50 años de su muerte, como uno de los grandes de su siglo',
'labels': ['sociedad', 'cultura', 'salud', 'economia', 'deportes'],
'scores': [0.3711881935596466,
0.25650349259376526,
0.17355826497077942,
0.1641489565372467,
0.03460107371211052]}
"""
The hypothesis_template parameter is important and should be in Spanish. In the widget on the right, this parameter is set to its default value: "This example is {}.", so different results are expected.
Metrics
| Model | Params | XNLI (acc) | *MLSUM (acc) |
|---|---|---|---|
| zs BETO | 110M | 0.799 | 0.530 |
| zs SELECTRA medium | 41M | 0.807 | 0.589 |
| zs SELECTRA small | 22M | 0.795 | 0.446 |
*evaluated with zero-shot learning (ZSL)
- XNLI: The stated accuracy refers to the test portion of the XNLI dataset, after finetuning the model on the training portion.
- MLSUM: For this accuracy we take the test set of the MLSUM dataset and classify the summaries of 5 selected labels. For details, check out our evaluation notebook
Training
Check out our training notebook for all the details.
