Version Control for Models, Data and Training Processes

Overview
In the PMI Certified Professional in Managing AI (PMI-CPMAI) framework, Version Control for Models, Data and Training Processes sits within the domain of supporting responsible and trustworthy AI. It is the disciplined practice of tracking, labeling, storing and governing every artifact that shapes an AI system. These artifacts include datasets, data transformations, feature definitions, code, hyperparameters, training environments, model binaries and evaluation results. The goal is to make any model reproducible, auditable, explainable and recoverable at any point in its lifecycle.

Why It Is Important
1. Reproducibility: AI outputs depend on the combination of data, code and configuration. Without versioning, a team cannot recreate a model's results. This undermines scientific validity and stakeholder trust.
2. Accountability and Auditability: Regulators, auditors and internal governance boards may ask which data trained a model, who approved it and when it was deployed. Version control provides the evidence trail, also called lineage or provenance.
3. Transparency and Explainability: To explain why a model behaves a certain way, you need to know exactly what went into it.
4. Bias and Fairness Investigation: If a deployed model produces biased outcomes, teams must trace the cause back to specific data versions, labeling changes or feature choices.
5. Risk Management and Rollback: When a new model degrades performance or causes harm, versioning allows a quick, controlled rollback to a known-good version.
6. Compliance: Frameworks such as the EU AI Act, GDPR, the NIST AI RMF and ISO/IEC 42001 emphasize documentation, traceability and record-keeping. GDPR adds obligations such as the right to erasure, which require knowing which data versions contain which records.
7. Collaboration: Multiple data scientists and engineers can work in parallel without overwriting each other's work.
8. Drift Management: Comparing current data and models against versioned baselines helps detect data drift and concept drift.

What It Is: The Key Components
Data Versioning: Capturing snapshots or immutable references of raw data, cleaned data, labeled data, training, validation and test splits, and feature sets. It includes metadata such as source, collection date, consent status, schema and quality metrics.
Model Versioning: Assigning unique identifiers to trained models, typically with semantic versioning (for example v2.1.0). Each version is stored with its weights, architecture, performance metrics and approval status, usually in a model registry.
Code Versioning: Using tools such as Git to manage training scripts, preprocessing pipelines and inference code.
Training Process and Experiment Versioning: Recording hyperparameters, random seeds, library versions, hardware, environment containers, training duration and evaluation outcomes for each experiment run.
Environment Versioning: Capturing dependencies through containers (for example Docker images) and requirements files so the runtime can be recreated.
Lineage and Provenance: The linked chain showing which data version and which code version, under which configuration, produced which model version, and where that model is deployed.
Documentation Artifacts: Model cards, datasheets for datasets and change logs that accompany each version.

How It Works
Step 1: Establish a Versioning Policy. Define naming conventions, what must be versioned, retention periods, access controls and approval gates. Align the policy with organizational AI governance.
Step 2: Version the Data. Use tools such as DVC, LakeFS, Delta Lake or Pachyderm to create immutable snapshots or hashes of datasets. Record metadata on source, consent, labeling guidelines version and quality checks.
Step 3: Version the Code. Commit all pipeline code to a repository with branching strategies, pull requests and code reviews.
Step 4: Track Experiments. Use tools such as MLflow, Weights and Biases, Neptune or Comet to log parameters, metrics and artifacts for every training run automatically.
Step 5: Register Models. Store promoted models in a model registry with stages such as Staging, Production and Archived. Attach metrics, fairness evaluations, approvals and model cards.
Step 6: Link Everything. Make sure each model version references its exact data version, code commit, environment and configuration. This creates end-to-end lineage.
Step 7: Integrate with CI/CD/CT (MLOps). Automate testing, validation, bias checks and deployment, with continuous training pipelines that create new versions under controlled conditions.
Step 8: Monitor and Govern. Compare production behavior with versioned baselines. Trigger retraining or rollback when thresholds are breached. Keep an audit trail of every change and approval.
Step 9: Retire and Archive. Decommission old versions according to retention policy while preserving records needed for audits or legal purposes.

Connection to the CPMAI Methodology
CPMAI follows iterative phases:

• Business Understanding
• Data Understanding
• Data Preparation
• Model Development
• Model Evaluation
• Model Operationalization

Version control supports every phase:

• In Data Understanding and Data Preparation, it captures data sources and transformations.
• In Model Development, it tracks experiments.
• In Model Evaluation, it ties metrics to specific versions.
• In Operationalization, it enables controlled deployment, monitoring and rollback.

Because CPMAI is iterative, versioning ensures each iteration is traceable and comparable to previous ones.

Common Challenges

• Large datasets that are expensive to snapshot. These are handled with hashing, deltas or pointer-based versioning.
• Unstructured data such as images, audio and text.
• Non-determinism in training. This is mitigated by fixing seeds and logging environments.
• Privacy obligations conflicting with immutability. This requires processes for data deletion that preserve audit records.
• Cultural resistance or ad-hoc experimentation habits among data scientists.

Exam Tips: Answering Questions on Version Control for Models, Data and Training Processes
Tip 1: Think beyond code. The exam often tests whether you know that versioning in AI covers data, models, hyperparameters, environments and training processes, not just source code. If an answer choice versions only the code, it is usually incomplete.
Tip 2: Link versioning to trustworthiness. When a question mentions auditability, transparency, accountability, reproducibility or regulatory compliance, version control and lineage are strong candidate answers.
Tip 3: Choose rollback for production failures. If a newly deployed model degrades or shows bias, the best immediate action is often to roll back to the previous approved version, then investigate using versioned lineage. Retraining from scratch is usually not the first step.
Tip 4: Reproducibility questions. If a team cannot reproduce results, the root cause is typically missing data snapshots, unlogged hyperparameters, unfixed random seeds or undocumented environments. Choose answers that introduce experiment tracking and data versioning.
Tip 5: Prefer proactive, systematic solutions. PMI exams favor establishing policies, processes and governance early, ideally starting in the Data Understanding phase, over reactive fixes. Select answers that build versioning into the pipeline from the start.
Tip 6: Recognize key terms. Know these terms:

• Model registry: a central store of model versions and their stages.
• Data lineage: where data came from and how it was transformed.
• Provenance: the origin and ownership of data.
• Experiment tracking: logging of training runs.
• Model card and datasheet: documentation that accompanies each version.

Tip 7: Bias investigations. When asked how to determine why bias appeared, the answer usually involves tracing back through versioned datasets and labeling changes to identify which change introduced the issue.
Tip 8: Governance and approvals. Expect scenarios where models move to production without sign-off. The correct answer typically involves approval gates within the model registry and documented change management.
Tip 9: Watch for distractors. Be wary of options such as:

• "Store only the final model."
• "Overwrite the dataset with cleaned data."
• "Rely on team memory or email."
• "Version only when regulators ask."

These violate traceability principles.
Tip 10: Balance with privacy. If a question involves a data deletion request under GDPR, the right answer acknowledges both honoring deletion and maintaining appropriate records. Versioning helps identify which models were trained on the affected data and may need retraining.
Tip 11: Iterative mindset. CPMAI is iterative and agile. Versioning enables safe iteration, so answers that support controlled experimentation and comparison across iterations align with the methodology.
Tip 12: Read the scenario's phase. Identify which CPMAI phase the scenario describes. In data phases, emphasize data versioning and lineage. In modeling phases, emphasize experiment tracking. In operationalization, emphasize the registry, deployment versioning, monitoring and rollback.

Sample Question Approach
Scenario: A credit-scoring model in production starts denying more applicants from a specific region after a recent update. Regulators request an explanation.
Best approach:

• Roll back to the last approved model version to limit harm.
• Use the model registry and lineage records to compare data versions, feature changes and hyperparameters between the two versions.
• Document the findings for regulators.
• Strengthen approval gates and fairness checks before future promotions.

Summary
Version control for models, data and training processes is a foundation of responsible and trustworthy AI. It turns AI development from an opaque, ad-hoc activity into a traceable, governable and reproducible discipline. On the exam, favor answers that are comprehensive (covering all artifacts), proactive (built in from the start), governance-oriented (approvals and audit trails) and risk-aware (rollback and monitoring).