Model Card for DeBERTa-v3-base-tasksource-nli

This is DeBERTa-v3-base fine-tuned with multi-task learning on 600 tasks of the tasksource collection. This checkpoint has strong zero-shot validation performance on many tasks (e.g. 70% on WNLI), and can be used for:

• Zero-shot entailment-based classification pipeline (similar to bart-mnli), see [ZS].
• Natural language inference, and many other tasks with tasksource-adapters, see [TA]
• Further fine-tuning with a new task (classification, token classification or multiple-choice).

[ZS] Zero-shot classification pipeline

from transformers import pipeline
classifier = pipeline("zero-shot-classification",model="sileod/deberta-v3-base-tasksource-nli")

text = "one day I will see the world"
candidate_labels = ['travel', 'cooking', 'dancing']
classifier(text, candidate_labels)

NLI training data of this model includes label-nli, a NLI dataset specially constructed to improve this kind of zero-shot classification.

[TA] Tasksource-adapters: 1 line access to hundreds of tasks

!pip install tasknet tasksource
import tasknet as tn
pipe = tn.load_pipeline('sileod/deberta-v3-base-tasksource-nli','glue/sst2') # works for 500+ tasksource tasks
pipe(['That movie was great !', 'Awful movie.'])
# [{'label': 'positive', 'score': 0.9956}, {'label': 'negative', 'score': 0.9967}]

The list of tasks is available in model config.json. This is more efficient than ZS since it requires only one forward pass per example, but it is less flexible.

Evaluation

This model ranked 1st among all models with the microsoft/deberta-v3-base architecture according to the IBM model recycling evaluation. https://ibm.github.io/model-recycling/

Software and training details

The model was trained on 600 tasks for 200k steps with a batch size of 384 and a peak learning rate of 2e-5. Training took 12 days on Nvidia A30 24GB gpu. This is the shared model with the MNLI classifier on top. Each task had a specific CLS embedding, which is dropped 10% of the time to facilitate model use without it. All multiple-choice model used the same classification layers. For classification tasks, models shared weights if their labels matched.

https://github.com/sileod/tasksource/
https://github.com/sileod/tasknet/
Training code: https://colab.research.google.com/drive/1iB4Oxl9_B5W3ZDzXoWJN-olUbqLBxgQS?usp=sharing

Citation

More details on this article:

@article{sileo2023tasksource,
  title={tasksource: Structured Dataset Preprocessing Annotations for Frictionless Extreme Multi-Task Learning and Evaluation},
  author={Sileo, Damien},
  url= {https://arxiv.org/abs/2301.05948},
  journal={arXiv preprint arXiv:2301.05948},
  year={2023}
}

Model Card Contact

damien.sileo@inria.fr

</details>