STAV: System Trustworthiness and Accountability Vocabulary

This work is still under development.

Ease the uses of standardised vocabularies in Python workflow.

stav:

• Norwegian: to spell (words)
• Swedish: letter (alphabet)
• Slovak: state (condition)
• Czech: stance

STAV taxonomy

Apart from providing access to standardised vocabularies, STAV also provide a minimal taxonomy at https://w3id.org/stav

STAV Python module

STAV terms can also be accessible through a Python module called stav, available freely on the Python Package Index.

STAV class names are accessible through constants in the stav module. These class names are in capital letters with underscores separating words, LIKE_THIS, as stated in PEP 8. Values of these constants are simply a STAV class name, a string in CamelCase.

For example, stav.INSTRUCTIONS_FOR_USE is a string with value of "InstructionsForUse". (In the future, it should be able to configure the casing to "instructions_of_use", etc.)

With this, it will make the standardization of documentation within an organization, or across organizations, easier and can facilitate the use of the terms in MLOps settings, where data scientists and data engineers can use STAV terms as keys in their model logging and registration.

Installation

pip install stav

Use with MLflow

import mlflow
from stav.vocab import dpv, spdx

with mlflow.start_run():
    mlflow.set_tag("data_subject_type", dpv.NaturalPerson)
    mlflow.set_tag("dataset_type",      spdx.dataset.DatasetType.text)
    mlflow.set_tag("confidentiality",   spdx.dataset.ConfidentialityLevelType.amber)

All terms are plain strings (IRIs), so they work anywhere a string is accepted. Use stav.label() to get a human-readable name from any term.

Use with Pitloom

from pitloom import loom
from stav.vocab import dpv, spdx

@loom.shoot("training_sbom.json")
def train(dataset_path):
    loom.set_model("my-model", model_type=dpv.ai.AITechnique.SupervisedLearning)
    loom.add_dataset(dataset_path, dataset_type=spdx.dataset.DatasetType.text)
    ...

Sister projects

STAV (a Vocabulary) and STAP (an ODRL Profile) are sisters for system trustworthiness and accountability.

• STAV provides a vocabulary (with focus on informational items) extracted from regulations and policy documents, mostly AI safety-related, like EU Artificial Intelligence Act draft. Its IRI is https://w3id.org/stav. Its code repository is at https://github.com/bact/stav/.

• STAP provides a set of core accountability relationships, based on Open Digital Rights Language. They are trying not to be AI-specific. Its IRI is https://w3id.org/stap. Its code repository is at https://github.com/bact/stap/.

License

STAV: System Trustworthiness and Accountability Vocabulary by Arthit Suriyawongkul is licensed under CC BY 4.0

The stav Python module is distributed under the terms of the Apache License 2.0.

This work has emanated from research conducted with the financial support of Taighde Éireann – Research Ireland under Grant number 18/CRT/6224 (Research Ireland Centre for Research Training in Digitally-Enhanced Reality (d-real)) and with the organizational support from members of these research groups:

• Knowledge and Data Engineering Group, Trinity College Dublin
• Transparent Digital Governance strand, ADAPT Centre

Related works

Members of RegTech group at ADAPT Centre contribute to AI and data ontology projects below, and they are may be of your interest:

• Pitloom - SBOM (software bill of materials) generator. Its SBOM fragments generation works with STAV vocabs.
• Trustworthy AI Requirements Ontology (TAIR)
• AI Risk Ontology (AIRO)
• Vocabulary of AI Risks (VAIR)
• Data Privacy Vocabulary (DPV) - with AI extenstion