fine-tuning vs prompting
We fine-tuned a model for five marketing jobs and the untrained base won four of them. This hub is the practical record of when training is worth it, and when the knowledge belongs in a prompt instead.
The measured evidence for this topic lives in The build log.
LoRA fine-tuning is cheap enough to run on a laptop, so the hard part moves to data and measurement. Seven lessons from our marketing adapters, with numbers.
When to fine-tune an LLM, measured on five marketing tasks where the untrained base beat our fine-tune four times, and the rule that explains why.
What Meta, Google and TikTok are automating in advertising, what the announcements actually ship, and what that leaves for an independent model to check.
Open-weights and small language models for marketing work: what exists, what they cost to run, and how they compare on a public benchmark.
Hooks, angles, ad copy and creative analysis: how to tag what an ad is doing, and what a model can and cannot learn about why it works.
How to tell whether a marketing AI is any good: public benchmarks, blind judging, label quality, seed variance and the claims they do and do not support.
Safe zones, character limits, aspect ratios and ad policy: the rules that decide whether an ad renders correctly before anyone judges whether it is good.