The goal
Train a model to write marketing copy in your brand's specific voice — matching tone, vocabulary, and structure. The classic distillation use case for content teams.
Data collection
- Collect 200–500 examples of approved brand copy (blog posts, social, emails).
- Pair each with a "brief" (the request that produced it).
- Include both good examples AND rewrites of off-brand copy → on-brand copy. The contrasts teach the model what to avoid as much as what to produce.
Evaluation is hard
Style transfer is one of the harder evaluation problems. Three approaches:
- A/B blind test: have your marketing team rate base vs fine-tuned outputs without knowing which is which.
- Brand voice rubric: score on specific dimensions (tone accuracy, vocabulary match, structure compliance).
- LLM-as-judge with the brand guidelines as system prompt: have GPT-4o rate adherence to written guidelines.