#3667·flair

[Question]: Is there existing or planned support for Hugging Face datasets in Flair?

Author: amina-mardiyyahCreated Jun 3, 2025Updated Jun 3, 2025
Labelsquestion

Question

Hi!

Thanks for the great work on this library — it's been incredibly useful for my current project.

I wanted to ask if there's currently any support or plans to support the Hugging Face datasets library in Flair — particularly for tasks like NER, POS, or classification where datasets.DatasetDict is a common format.

Right now, it looks like Flair expects CoNLL-style .txt files stored somewhere in local directory(data_folder), and loads them via ColumnCorpus, which works fine, but introduces some boilerplate when working with datasets objects directly (e.g., from load_dataset("conll2003") or custom HF datasets).

A few questions:

  • Is there an existing wrapper or method I might have missed?
  • Would you be open to supporting Hugging Face datasets natively — perhaps via a from_hf_dataset(...) method or similar utility?

Many thanks again.