Streaming datasets fail with 403 SignatureError: invalid key pair id from xet-bridge-us (Colab, latest libraries)

[mahmoud-mannes]

Describe the bug

I'm unable to stream datasets from the Hugging Face Hub. The issue affects both public datasets (e.g. beans) and gated datasets (ILSVRC/imagenet-1k).

The failure occurs before any examples are yielded. The request is redirected to the Xet CDN (xet-bridge-us) and fails with:

403 Forbidden

Auth failed: SignatureError: invalid key pair id

This started around July 13–14, 2026. The same notebooks had been working previously.

Steps to reproduce the bug

from datasets import load_dataset

ds = load_dataset(
    "AI-Lab-Makerere/beans",
    split="train",
    streaming=True,
    token=True,
)

next(iter(ds))

fails immediately.

Likewise,

from datasets import load_dataset

dataset = load_dataset(
    "ILSVRC/imagenet-1k",
    split="train",
    streaming=True,
)

for i, x in enumerate(dataset):
    print(i)

fails immediately.

The request is made against https://us.gcp.cdn.hf.co/xet-bridge-us/... and returns:

403 Forbidden

Auth failed:
SignatureError: invalid key pair id

The traceback ends with:

``'
HfHubHTTPError:

403 Forbidden

Cannot access content at:
https://us.gcp.cdn.hf.co/xet-bridge-us/...

Auth failed:
SignatureError: invalid key pair id
'``

The URL also contains: Key-Pair-Id=01KXEF4KZ1B6FV465MAWR4M21F

Expected behavior

Streaming public datasets should work without returning

403 Forbidden

Auth failed:
SignatureError: invalid key pair id`
``

### Environment info

- `datasets` version: 5.0.0
- Platform: Linux-6.6.122+-x86_64-with-glibc2.35
- Python version: 3.12.13
- `huggingface_hub` version: 1.23.0
- PyArrow version: 25.0.0
- Pandas version: 2.2.2
- `fsspec` version: 2025.3.0

[lhoestq]

Hi ! I couldn't reproduce the issue on my side. I can see two options, let me know if any helps you:

1. this was a transient issue (we restarted some machines recently but this shouldn't have impacted you tbh), can you try again just in case ?
2. something is wrong with your token smh, can you logout and login again ?

Let me know if the issue persists

[mahmoud-mannes]

Thanks for looking into it!

I've already tried both of your suggestions:

• I restarted the Colab session multiple times with fresh runtimes, the error still persisted (started yesterday about 14 hours ago, still happening)
• Logged out and back into Hugging Face (huggingface-cli logout / login) and verified that my token works. I've even tried doing it without a token using public datasets to no avail.

Small downloads (like the README.md) do work, but any type of streaming doesn't. I've replicated this with both ILSVRC/imagenet-1k (using my token) and AI-Lab-Makerere/beans , both don't work.

If there's any additional logging or environment information that would help, I'm happy to provide it.

[Ar4l]

I am also running into the same issue, flakily depending on which redirect server is being chosen, from GCP VMs.
Also reported here: huggingface/xet-core#897

[mahmoud-mannes]

Have there been no work arounds? Do we just have to wait?

[Ar4l]

@mahmoud-mannes for me, HF_HUB_DISABLE_XET=0 seems to work; issue only occurs when it's redirected to XET and XET is disabled.

[espressosession]

Same error via diffusers on Colab, so this isn't specific to datasets: same Key-Pair-Id: 01KXEF4KZ1B6FV465MAWR4M21F, huggingface_hub 1.23.0, Python 3.12.13.

Two observations that might help narrow it down:

• It also hits public repos (stable-diffusion-v1-5/stable-diffusion-v1-5), on presigned URLs carrying user_id=public, and it happens with and without a token.
• It is intermittent: the same file 403s, then downloads fine minutes later.

HF_HUB_DISABLE_XET=1 does not avoid it — you are still redirected to xet-bridge-us — but you get the 403 as an exception instead of a download that stalls part-way and never times out. That made a retry loop possible, and since the cache resumes, it does eventually get through:

import os
os.environ["HF_HUB_DISABLE_XET"] = "1"   # before huggingface_hub is imported

import time

def fetch(call, retries=40):
    for attempt in range(1, retries + 1):
        try:
            return call()
        except Exception as exc:
            if "SignatureError" not in str(exc) or attempt == retries:
                raise
            time.sleep(3)

Not a fix, but it unblocked a multi-GB checkpoint download for me.

[Akrao9]

Same error on Colab while streaming:

• 403 Forbidden from https://us.gcp.cdn.hf.co/xet-bridge-us/...
• Auth failed: SignatureError: invalid key pair id
• Key-Pair-Id=01KXEF4KZ1B6FV465MAWR4M21F

Looks like the same GCP CDN signing issue as huggingface/xet-core#897.

[lhoestq]

Hi ! Thanks we are in the process of restarting our CDN for gcp / google colab. It should be back soon, sorry for the inconvenience

[agiveon]

Seems to work again

[Akrao9]

Yes it’s working

[Favioleiva]

Same issue here with hf_hub_download on a private dataset repo, from Google Colab.

• Requests are redirected to https://us.gcp.cdn.hf.co/xet-bridge-us/...
  and fail with 403 Forbidden — Auth failed: SignatureError: invalid key pair id
• Same Key-Pair-Id=01KXEF4KZ1B6FV465MAWR4M21F as the OP
• Started around July 13–14, 2026. Same notebook had been working fine before with no code changes
• Still reproducible as of July 15, 2026
• Request IDs: 01KXFP39MES5SH6ZTBR3TC8CFS, 01KXFP3KXQFG20PPBF1QVZBQBQ, 01KXFP47YXK7824D1V9BS6TFRH, 01KXFQ1EKJMK4MEP49RC0QW9TN
• huggingface_hub: 1.20.1
• hf-xet: 1.5.1
• Python 3.12, Google Colab