# Tasksource - Datasets harmonization for frictionless multi-task/evaluation

**URL:** <https://discuss.huggingface.co/t/tasksource-datasets-harmonization-for-frictionless-multi-task-evaluation/31442>\
**Category:** 🤗Datasets\
**Created:** [February 10, 2023, 9:41am UTC](https://discuss.huggingface.co/t/tasksource-datasets-harmonization-for-frictionless-multi-task-evaluation/31442 "2023-02-10T09:41:16Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![sileod](https://avatars.discourse-cdn.com/v4/letter/s/f19dbf/32.png) [@sileod](https://discuss.huggingface.co/u/sileod)\
**Post date:** [February 10, 2023, 9:41am UTC](https://discuss.huggingface.co/t/tasksource-datasets-harmonization-for-frictionless-multi-task-evaluation/31442/1 "2023-02-10T09:41:16Z")

</div>

Hi everyone, I’m the author tasksource. I was annoyed dataset alignment/harmonization for multitask learning, so I created a concise “[language](https://github.com/sileod/tasksource/)” to represent the harmonization preprocessings.

With a pip install and 3 lines of code, you can load hundreds of datasets that can be used interchangeably. I also trained a [deberta-base-tasksource](https://huggingface.co/sileod/deberta-v3-base-tasksource-nli) multitask model on all tasksource tasks and obtained very good results, I would say that it’s the best “base-size” model on huggingface for classification/NLI/multiplechoice (it beats all of them on model-recycling evaluation.

I’m open to contributions and suggestions, thanks !