数据集:

tasksource/zero-shot-label-nli

英文

tasksource 通过自然语言推理重新定义了分类任务。该数据集旨在提高对 zero-shot classification HF pipelines 的标签理解。

输入的文本对由换行符 (\n) 分隔。

from transformers import pipeline
classifier = pipeline(model="sileod/deberta-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"],
)

现在, deberta-v3-base-tasksource-nli 的训练组合中包含了 label-nli(只占很小一部分,以保持模型的通用性,但请注意,nli模型可以在没有具体监督的情况下对类似标签的零-shot分类起作用)。

@article{sileo2023tasksource,
  title={tasksource: A Dataset Harmonization Framework for Streamlined NLP Multi-Task Learning and Evaluation},
  author={Sileo, Damien},
  year={2023}
}