`huggingface_ner` ML Backend: fit() fails with KeyError 'annotation' for certain webhook payloads

[yuka-with-data]

Description

When using the huggingface_ner ML Backend in model.py, the fit() method can fail during webhook-triggered training due to an assumption about the webhook payload structure.

Problem

The current implementation assumes that the webhook payload always contains:

data['annotation']['project']

However, in some cases (e.g., after submitting token-level annotations), the payload does not include the annotation key, causing the backend to crash.

Also, according to the webhook payload details in the documentation, project is a top-level field and not inside annotation, so assuming data['annotation']['project'] can lead to a KeyError.

When submitting annotations, the following error occurs:

Traceback (most recent call last):
  File ".../label_studio_ml/api.py", line 126, in webhook
    result = model.fit(event, data)
  File ".../huggingface_ner/model.py", line 146, in fit
    project_id = data['annotation']['project']
KeyError: 'annotation'

Environment

• Setup: Local (non-Docker) ML Backend
• Model: huggingface_ner example
• Trigger: Submitting token-level annotations

Suggested Fix

Handle webhook payloads more defensively by avoiding assumptions about the presence of the annotation key when extracting project_id.

I'd be happy to open a PR implementing this change and linking it to this issue.