"no-token" is passed to `huggingface_hub` when token is `None`

[Wauplin]

Describe the bug

In the 2 lines listed below, a token is passed to huggingface_hub to get information from a dataset. If no token is provided, a "no-token" string is passed. What is the purpose of it ? If no real, I would prefer if the None value could be sent directly to be handle by huggingface_hub. I feel that here it is working because we assume the token will never be validated.

datasets/src/datasets/load.py

Line 753 in 5b23f58

token=token if token else "no-token",

datasets/src/datasets/load.py

Line 1121 in 5b23f58

token=token if token else "no-token",

Expected results

Pass token=None to huggingface_hub.

Actual results

token="no-token" is passed.

Environment info

huggingface_hub v0.10.0dev

[albertvillanova]

Hi @Wauplin, thanks for raising this potential issue.

The choice of passing "no-token" instead of None was made in this PR:

• Properly raise FileNotFound even if the dataset is private #4536

According to the PR description, the reason why it is passed is to avoid that HfApi.dataset_info uses the local token when no token should be used.

[Wauplin]

Hi @albertvillanova , thanks for finding the original issue 👍

As of next release of huggingface_hub, the token argument will be deprecated in favor of the use_auth_token argument in dataset_info method. This change as been done by @SBrandeis in huggingface/huggingface_hub#928. use_auth_token is a bit different and allow the case "don't sent the cached token by default".

If you want to strictly avoid sending the cached token from datasets, you can use:

# token=token if token else "no-token", <- will fail because token is not valid

use_auth_token=token if token else False,  # using the new `use_auth_token` parameter

And as a note, I am currently updating the "don't send the cached token by default"-rule to "don't send the cached token on public repos by default but use it in private ones" in huggingface/huggingface_hub#1064. This will not change the fact that use_auth_token=False doesn't send the token at all.

[Wauplin]

What is current strategy in term of updating huggingface_hub version in datasets ? I don't want to break stuff in the next release so let's find a proper solution :)

[lhoestq]

As soon as token is deprecated and hfh has a new release, we'll update datasets to use the new argument instead. Does it sound good to you ?

[Wauplin]

Perfect 👌

[albertvillanova]

Hi @Wauplin, thanks for the warning about the deprecation of token in favor of use_auth_token.

Indeed, in datasets we use internally use_auth_token, which in this case was transformed to token to call HfApi.dataset_info:

datasets/src/datasets/load.py

Line 747 in 1a9385d

token = HfFolder.get_token() if self.download_config.use_auth_token else None

Therefore, for the new hfh release, the fix will be trivial: we will pass directly use_auth_token.

As discussed during our meeting yesterday, due to the fact that at datasets we support multiple hfh versions, I think we should handle passing token or use_auth_token depending on the hfh version.