cerbia-ml | CerbIA

cerbia-ml provides cerbia.ml, an optional runtime layer for pinned model artifacts and text-classification adapters. Installing it alone does not add a CLI command or a scanner to a core configuration.

Install

pip install cerbia-ml

The distribution depends on cerbia-core and the ONNX Runtime support provided through Optimum.

Resolve model artifacts

ArtifactCoords identifies an artifact by repository reference, a required 40-character revision, and optional subfolder or filename. The ArtifactSource protocol resolves coordinates to a local path and reports whether an artifact is already available. HuggingFaceHubSource checks the Hugging Face cache and can fetch a pinned snapshot, limiting the download to a requested subfolder or file when those coordinates are set.

from cerbia.ml.artifacts import HuggingFaceHubSource
from cerbia.ml.models import ArtifactCoords

artifact = ArtifactCoords(
    ref="ORG/MODEL",
    revision="_40_CHARACTER_COMMIT_",
    filename="model.onnx",
)
source = HuggingFaceHubSource()
if not source.is_available(artifact):
    local_path = source.fetch(artifact)

Replace ORG/MODEL with the repository and 40_CHARACTER_COMMIT with its immutable commit. The revision must contain exactly 40 lowercase hexadecimal characters.

Classify text

HuggingFaceClassifierAdapter loads a tokenizer and an ONNX sequence classification model, then returns validated results for a batch of strings. Its constructor accepts model and tokenizer artifact coordinates, an adapter configuration, and local_files_only, which defaults to true.

from cerbia.ml.adapters.classifiers.huggingface_classifier import (
    HuggingFaceClassifierAdapter,
    HuggingFaceClassifierAdapterConfig,
)

With local_files_only: true, the required model and tokenizer assets must already be cached. Set it to false only when runtime downloads are suitable for the deployment. The optional cerbia-protectai integration uses this package for model and tokenizer loading. See the product-wide configuration reference and quickstart for pipeline setup.