What Is AI Fine-Tuning?

Part 8 established fine-tuning's real, correct use case — genuinely changing model behavior, not injecting current facts. This part covers the actual, technical mechanism, building directly on the AI Fundamentals series' own neural network training explanation.

A real, precise definition

Fine-tuning: taking an already-trained model and continuing real
  training on it with a smaller, genuinely task-specific real
  dataset, adjusting its existing parameters further rather than
  training a new model from scratch

This directly builds on the AI Fundamentals series' own pretraining coverage — fine-tuning starts from a model that's already broadly capable (having completed the massive, real pretraining phase), and does a real, smaller, second phase of training specifically shaping it toward a narrower, desired real behavior.

The real, required training data format

[
  {
    "messages": [
      {"role": "user", "content": "Draft a follow-up email for a lead who requested a demo."},
      {"role": "assistant", "content": "Subject: Following up on your GreenDesk demo\n\nHi [Name],\n\nThanks for..."}
    ]
  },
  {
    "messages": [
      {"role": "user", "content": "Draft a follow-up email for a lead who went quiet after pricing."},
      {"role": "assistant", "content": "Subject: Checking in\n\nHi [Name],\n\nWanted to..."}
    ]
  }
]

Real fine-tuning data is genuinely just many, real examples of the exact input-output pattern you want the model to learn — for GreenDesk's real email-formatting need from part 8, this means real, actual examples of correctly-formatted emails, in the exact real format the legal team requires, repeated across enough real, varied examples for the model to learn the underlying pattern rather than memorize specific instances.

What genuinely happens to the model's real parameters

This is exactly the AI Fundamentals series' own training mechanism,
  applied again: real gradient descent, adjusting the model's real,
  existing weights slightly with each real training example, so its
  predictions increasingly match the fine-tuning dataset's actual
  patterns

The real, key technical difference from pretraining is scale — fine-tuning uses a genuinely much smaller dataset and a genuinely much shorter training run, since the goal is nudging existing, real capability toward a specific pattern, not building general language capability from nothing.

Full fine-tuning vs. LoRA: a real, practical distinction

Full fine-tuning: adjusts ALL of the model's real parameters —
  genuinely expensive, requiring substantial real compute
LoRA (Low-Rank Adaptation): a real, practical technique that adds a
  small number of NEW, trainable parameters alongside the original,
  frozen ones — achieving similar real, practical results at
  dramatically lower real cost and compute

LoRA is the real, more commonly used approach for most practical fine-tuning today — genuinely, it works by learning a small, real set of adjustments layered on top of the existing, unchanged model weights, rather than modifying the massive, original parameter set directly, which is why it's dramatically cheaper while often achieving genuinely comparable results for a well-scoped, real task.

A real, honest, minimum viable dataset size

Real, practical guidance varies significantly by provider and task
  complexity, but genuinely useful fine-tuning results typically
  require at minimum dozens to hundreds of real, high-quality
  examples — not thousands, but also not just five or ten

This real number matters directly for a practical decision: GreenDesk's real email-formatting fine-tune (part 8's example) is realistic if the team can genuinely produce a hundred-plus real, correctly-formatted example emails; it's a real, premature undertaking if only a handful of examples currently exist.

What fine-tuning does NOT do

It does NOT give the model access to real, current information the
  way RAG does (part 8's core distinction) — fine-tuning data
  becomes frozen at training time, exactly the same real staleness
  limitation covered throughout the AI Fundamentals series

This is worth restating precisely, given how often it's confused: fine-tuning teaches behavior patterns, not current facts — a model fine-tuned on real email examples from six months ago doesn't automatically know today's actual GreenDesk pricing any more than the unmodified base model would.

Next: when fine-tuning actually makes sense — a real, concrete decision framework and honest cost accounting before committing to it.