Data Curation

Data curation makes data fit for a specific purpose rather than merely collecting or storing it. For foundation model fine-tuning, that purpose might be teaching a model to follow a company’s response format, classify support requests, or answer questions in an appropriate style. Curation starts by defining the desired behavior and identifying which examples demonstrate it. A large collection of documents is not automatically a useful fine-tuning dataset.

Engineers first assess source quality, ownership, licensing, consent, and permitted uses. Publicly accessible material is not necessarily licensed for model training, and internal access does not automatically authorize reuse. Sensitive fields such as customer identifiers, credentials, and health information require appropriate removal, masking, or other approved controls. Dataset metadata should record provenance, transformations, reviewers, and usage restrictions so that examples remain traceable.

Preparation then includes removing corrupted records, resolving inconsistent formats, detecting duplicates, and validating examples against the selected model’s requirements. Supervised fine-tuning commonly uses input-output examples or conversations containing approved responses. Exact fields, supported roles, file formats, token limits, and dataset size requirements depend on the model and training platform. Truncating a conversation without checking its meaning can remove the information needed to justify the expected answer.

Representativeness means covering the tasks, users, languages, difficulty levels, and operating conditions the model will encounter. For a support assistant, a dataset containing only successful password resets will not teach reliable handling of billing disputes or ambiguous requests. Include appropriate examples of clarification, escalation, and refusal, not just straightforward answers. Rare but important situations may need deliberate sampling, while evaluation should also measure performance against the expected production distribution.

Labels and target responses need clear annotation instructions and quality review. Subject-matter experts can resolve ambiguous cases, while reviewer agreement checks reveal unclear guidance. In reinforcement learning from human feedback, reviewers may compare or rank model responses rather than supply a single correct answer. In a common RLHF workflow, these preferences train a reward model that guides further optimization. Preference data therefore needs consistent criteria for helpfulness, correctness, and safety; collecting feedback alone is not RLHF.

Dataset size is a trade-off, not a universal target. Too little data may miss important behaviors or encourage overfitting, while more low-quality data can reinforce errors and increase labeling and training costs. Start with a carefully reviewed dataset, evaluate a baseline and the fine-tuned model, and add examples that address measured weaknesses. Fine-tuning is also not a reliable substitute for retrieving frequently changing facts from an authoritative source.

In the cloud, teams commonly store source data and approved dataset versions separately in object storage, run repeatable transformation jobs, and restrict access through identity and data permissions. They retain version identifiers, dataset manifests, and validation results to connect a model artifact to its training inputs. Training, validation, and test sets should be separated before training, with checks that duplicates or closely related records do not cross boundaries. Grouping related conversations or documents during splitting helps prevent leakage and misleading evaluation scores.

Curation continues after deployment as error analysis identifies missing scenarios and source policies change. Automation makes validation and filtering scalable, but human review remains important for subtle correctness and bias issues. Cleaning is only one part of curation: neither an error-free file nor a larger dataset guarantees representative coverage, lawful use, or useful model behavior.