Datasets:
metadata
dataset_info:
features:
- name: labels
dtype:
class_label:
names:
'0': entailment
'1': neutral
'2': contradiction
- name: premise
dtype: string
- name: hypothesis
dtype: string
- name: task
dtype: string
splits:
- name: train
num_bytes: 185352754
num_examples: 878967
- name: test
num_bytes: 1775890
num_examples: 9400
- name: validation
num_bytes: 1817480
num_examples: 9400
download_size: 104413879
dataset_size: 188946124
license: other
task_categories:
- zero-shot-classification
- text-classification
task_ids:
- natural-language-inference
multilinguality:
- multilingual
mtasksource classification tasks recasted as natural language inference. This dataset is intended to improve label understanding in zero-shot classification HF pipelines.
Inputs that are text pairs are separated by a newline (\n).
from transformers import pipeline
classifier = pipeline(model="sileod/mdeberta-v3-base-tasksource-nli")
classifier(
"I have a problem with my iphone that needs to be resolved asap!!",
candidate_labels=["urgent", "not urgent", "phone", "tablet", "computer"],
)
mdeberta-v3-base-tasksource-nli will include label-nli in its training mix (a relatively small portion, to keep the model general, but note that nli models work for label-like zero shot classification without specific supervision (https://aclanthology.org/D19-1404.pdf).
@article{sileo2023tasksource,
title={tasksource: A Dataset Harmonization Framework for Streamlined NLP Multi-Task Learning and Evaluation},
author={Sileo, Damien},
year={2023}
}