Decide when to fine-tune

Choose the kind of change you need

Prompting

Improve the task description, provide a few examples, or define the output format. Changes take effect on the next request.

Retrieval

Fetch up-to-date reference material and supply it at inference time. Use this when facts change or answers need evidence.

Fine-tuning

Train on examples to change how the model behaves across inputs. LoRA trains a smaller set of parameters while keeping the original model weights frozen.

Prepare an experiment

1. Choose one stable behavior, such as classifying fictional library tickets into four categories.
2. Collect labeled examples you have permission to use. Remove duplicates before dividing them into training, validation, and held-out test sets.
3. Measure a prompt-only baseline on the held-out set. Record model revision, prompt, and criteria before looking at failures.
4. Choose a supported training stack and a matching base model. Record training settings and keep the original model for comparison. On Apple silicon, MLX LM documents supported LoRA workflows.
5. Compare both variants on the same held-out data, including rare cases and new phrasing. Check for regressions outside the target behavior before replacing the baseline.

Try it without training

Create ten fictional tickets with four categories. Use five to write clear class definitions and examples. Keep the rest for a first prompt-only check. If failures cluster around unclear labels, clarify the rules and collect better examples before training. Use a larger held-out set to evaluate an actual adapter.

Check your understanding

A handbook changes every week and employees need answers with citations. Which technique should you try first?

Use retrieval with the current handbook and verify the citations. Use fine-tuning for a stable behavior only if the baseline leaves a measurable gap.

Sources

• Hugging Face LLM Course · Fine-tuning
• LoRA · Original paper
• MLX LM · LoRA training guide