Mise au point

It changes learned parameters to adapt behavior. It differs from adding examples to a prompt or retrieving documents at answer time, and it does not automatically keep factual information current.

Points clés à retenir

• Define the behavior to adapt.
• Compare simpler alternatives.
• Evaluate gains and regressions on held-out tasks.

Define the behavior that needs to change. Consistent output style, a specialized classification task, and use of recent facts are different requirements. Prompting or retrieval may solve some of them without a training job. Compare those alternatives before adding model-maintenance work. Build examples that reflect the intended behavior and include difficult cases. Keep a held-out evaluation set separate from training and tuning decisions. Review labels, duplicate records, permissions, and any confidential information before using the dataset. Adaptation can update all parameters or a selected subset, depending on the method. Lower memory or fewer trainable parameters do not eliminate the need to evaluate the resulting model. Check both the target task and capabilities that should remain intact. Record the base model, data version, training settings, and resulting checkpoint. Evaluate deployment costs, response time, and rollback before release. When the source knowledge changes, decide whether to update retrieval, revise the dataset, retrain, or change the product’s evidence workflow.

1. Imagine a support assistant that knows how to answer clearly but needs a policy updated every week.
2. Start by testing retrieval of the current policy rather than retraining merely to insert the latest wording.
3. If the actual problem is persistent failure to follow a stable response format, compare prompt changes and a carefully evaluated fine-tuning dataset.

This constructed decision separates changing evidence from changing learned behavior.