TableDetector() crashes with transformers >= 5.x: "Field 'dilation' expected bool, got NoneType"

[ofermend]

Summary

With transformers 5.x installed, constructing TableDetector() (or AutoTableFormatter()) raises immediately, making gmft unusable:

File ".../gmft/detectors/tatr.py", line 48, in __init__
    self.detector = TableTransformerForObjectDetection.from_pretrained(
...
File ".../huggingface_hub/dataclasses.py", line 146, in __strict_setattr__
    raise StrictDataclassFieldValidationError(field=name, cause=e) from e
huggingface_hub.errors.StrictDataclassFieldValidationError: Validation error for field 'dilation':
    TypeError: Field 'dilation' expected bool, got NoneType (value: None)

Reproduction

from gmft.auto import TableDetector
TableDetector()  # raises

Versions

• gmft 0.4.2 and 0.4.3 (both affected)
• transformers 5.8.1 (also reproduced on 5.14.1)
• huggingface-hub 1.13.0
• Python 3.12
• Works fine with transformers 4.57.6

Root cause

transformers 5.x validates model configs with huggingface_hub's strict dataclasses. The microsoft/table-transformer-detection checkpoint's config.json on the Hub contains "dilation": null, which fails the bool field validation inside `from_pretcode runs. The structure-recognition checkpoint loads through the same path.

Why staying on 4.x is problematic

transformers < 5.3.0 / < 5.5.0 carry two high-severity advisories with no 4.x

• CVE-2026-4372 / GHSA-29pf-2h5f-8g72 (RCE via malicious model config.json)
• CVE-2026-5241 / GHSA-fgcw-684q-jj6r (arbitrary code execution during model

So projects that upgrade transformers for these fixes lose gmft.