Data governance defines who can make decisions about data and how those decisions are enforced and verified. It covers ownership, stewardship, acceptable use, quality, privacy, security, retention, and accountability. Data management performs activities such as storing, transforming, and deleting data; governance establishes the rules and responsibilities governing those activities. A catalog or access-control system supports governance, but neither replaces organizational policies and accountable owners.
In the cloud, governance commonly begins with an inventory of datasets, their owners, sensitivity classifications, source systems, and approved uses. Engineers translate policy into access permissions, encryption settings, approval workflows, retention controls, and audit records. For example, a customer-support dataset might be approved for operational reporting but not for model training. Access to that dataset does not automatically grant permission to reuse it for fine-tuning.
Preparing data to fine-tune a foundation model requires both governance decisions and technical curation. A team should verify licensing, consent where required, contractual restrictions, and whether the intended training use is permitted. It should identify sensitive information, remove unnecessary fields, and apply suitable redaction or de-identification. Masking names alone does not guarantee anonymity because combinations of remaining attributes can still identify individuals.
Curation should remove corrupted records, duplicates, irrelevant examples, and unsuitable content while preserving examples needed for the intended task. Dataset size must meet the selected model's requirements, but there is no universally correct number of examples. More data increases processing, labeling, and review costs and can worsen results when it introduces noise or bias. A smaller, carefully curated dataset can be more effective than a larger, poorly controlled collection.
Representativeness means covering the users, languages, tasks, and operating conditions expected in production, including important rare cases. Teams should document coverage gaps and evaluate performance across relevant groups. Training, validation, and test sets should be separated appropriately, often by customer, document, or time period, rather than merely by randomly splitting closely related records. This reduces leakage that would otherwise make evaluation results appear better than real-world performance.
For supervised fine-tuning, governance includes clear labeling instructions, qualified reviewers, disagreement resolution, and checks for consistent target responses. Reinforcement learning from human feedback, or RLHF, may use human preferences between model responses to train a reward model that guides subsequent optimization. These preference datasets need documented evaluation criteria, reviewer training, privacy protections, and checks for systematic bias. RLHF is not simply another name for labeling, and it is not required or supported by every fine-tuning workflow.
A practical support-assistant project might retain restricted source conversations separately from an approved, redacted training dataset. Each dataset version records its source, transformations, labeling guidelines, approvals, and associated training runs. This traceability supports investigation and reproducibility. Retention policies must also consider derived datasets and backups; deleting a source record does not automatically remove its influence from an already trained model.
Governance introduces review effort and can slow experimentation, especially when approvals are manual. Reusable policies, automated validation, and risk-based review help reduce that overhead. The goal is controlled, useful data access, not blanket restriction. Encryption and regulatory compliance are important, but neither alone establishes data quality, representativeness, or permission to train a model.