Apppublic

minotianadev/iso-42001-mlops-dashboard

sourceHugging Faceupdated 3mo agoView on Hugging Face

0likes

App README

Auditable MLOps Pipeline — ISO/IEC 42001 & 38505

An end-to-end ML pipeline that automates data validation, model training, bias mitigation, and metrics tracking, producing a signed compliance audit report (data/compliance_audit_report.json) every time a model is built.

Quick Start

bash

pip install -r requirements.txt
python -m src.train          # runs ingest -> train -> audit -> certificate
pytest -v                    # 13 tests: regression + compliance schema/tamper checks
streamlit run src/app.py     # launch the compliance dashboard

What gets generated

—data/model.joblib — trained RandomForestClassifier
—data/lineage_log.json — append-only ledger of every data ingestion event (ISO/IEC 38505)
—data/compliance_audit_report.json — signed certificate with dataset hash, model performance, and fairness metrics (ISO/IEC 42001 + 38505)

Architecture

Raw CSV -> ingest.py (SHA-256 hash, lineage log)
        -> train.py (RandomForestClassifier, accuracy/F1)
        -> audit.py (Disparate Impact Ratio, Demographic Parity Difference)
        -> certificate.py (signed JSON certificate)
        -> app.py (Streamlit dashboard, certificate download)

Fairness threshold

The Disparate Impact Ratio must fall within [0.80, 1.25] to be marked COMPLIANT. Outside that range the certificate is marked NON_COMPLIANT (or NON_COMPLIANT_RISK_OVERRIDE if explicitly overridden).

Deployment (free tier)

Push this repo to a Hugging Face Space (Streamlit SDK) or Render free web service. See SPECIFICATION.md for full deployment notes and the .github/workflows/compliance_ci.yml workflow for CI verification on push.