Fine-tuning and Adaptation roadmap
Changing a model's weights or behaviour to fit your problem. A real capability with a narrow band of correct application, surrounded on both sides by cheaper answers that people skip past.
Deep dive11 stages170 topics
Section 1 comes first deliberately. Most fine-tuning projects should not have started, and the ability to argue that clearly is worth more than the ability to run the training job. Work through it before anything else - if the answer is that you shouldn't fine-tune, you have saved months and the roadmap has done its job.
The Decision
The most valuable section here, and the one that saves the most time. Work it before you do anything else.
Foundations
Dataset Construction
Where the outcome is actually decided. A thousand mediocre examples lose to two hundred excellent ones, consistently.
Methods
Training
Preference Tuning
Aligning to preferences rather than to examples. Higher effort, narrower application - reach for it only when supervised tuning has plateaued.
Distillation
Evaluation
A tuned model you cannot evaluate is a liability. The comparison that matters is against the base model on data neither has seen.
Serving
Lifecycle
The part that turns a successful experiment into an ongoing cost. Plan it before you train, not after.
Tooling
Before you train anything
Answer these honestly. If you cannot answer all seven, you are not ready, and the training run will produce a model you cannot evaluate or maintain.
- What exactly is the base model getting wrong? Named failure modes with frequencies, from a real eval set - not an impression.
- Have you exhausted prompting and retrieval? Better instructions, better examples, better context. Most reported fine-tuning wins were prompt problems.
- Is this a knowledge problem or a behaviour problem? Knowledge belongs in retrieval. Fine-tuning teaches format, style and task shape - it teaches facts badly and expensively.
- Do you have the data? Hundreds to thousands of genuinely good examples, and a plan for who produces more when the model changes.
- How will you evaluate it against the base? A held-out set that existed before you started training.
- Who maintains it in a year? Base models get deprecated. Your tuned model does not follow automatically.
- What does serving it cost versus an API call? Including idle GPU time, not just tokens.
The strongest reasons to fine-tune are consistent output format at scale, a narrow repeated task where a small model can replace a large one, latency or cost that a smaller model fixes, and behaviour that resists instruction. Those are real. They are also less common than the enthusiasm suggests.
Fine-tuning moves the distribution of what the model tends to produce. It does not install facts, it does not add reasoning ability, and it will confidently make things up in your house style. Most disappointment in this area traces back to expecting the first of those and getting the third.