Hugging Face datasets "FileNotFoundError: Couldn't find a dataset script" in CI

load_dataset could not locate the dataset locally or on the Hub. The name is wrong or unqualified, a local path does not exist, or a private dataset was requested without a token so the Hub hides it.

What this error means

load_dataset fails with "FileNotFoundError: Couldn't find a dataset script at <path> or any data file in the directory" on the runner.

FileNotFoundError: Couldn't find a dataset script at /home/runner/work/app/my_dataset/my_dataset.py
or any data file in the directory. Couldn't find 'my_dataset' on the Hugging Face Hub either.

Common causes

The dataset id or local path is wrong

A typo, a missing org/ prefix, or a path that does not exist on the runner means datasets finds nothing to load.

A private dataset requested without a token

Without HF_TOKEN, the Hub hides the private dataset, so it appears missing.

How to fix it

Use the exact dataset id and set a token if private

1. Copy the dataset id from its Hub page (including the org).
2. For a local dataset, confirm the path exists in the checkout.
3. Set HF_TOKEN in the job for a private dataset.

env:
  HF_TOKEN: ${{ secrets.HF_TOKEN }}

Load with the correct config name

Some datasets require a config; pass it explicitly so datasets resolves the right files.

from datasets import load_dataset
load_dataset("glue", "mrpc")

How to prevent it

• Copy dataset ids and config names from the Hub page.
• Set HF_TOKEN for private datasets in CI.
• Verify local dataset paths exist after checkout.

Related guides