Fine-tuning in Practice
The difference between fine-tuning that works and fine-tuning that fails is almost always in the dataset. 1,000 high-quality examples beat 100,000 mediocre ones. LoRA reduces the compute bill by 99%. Careful evaluation catches catastrophic forgetting before deployment. And knowing when NOT to fine-tune saves weeks.
The form letter
Your team writes short, warm replies. The model keeps writing form letters.
You’ve rewritten the prompt for weeks and it still drifts. Is it time to fine-tune?
Prompt, look up, train
Fine-tuning isn’t the only fix. Which tool fits each problem?
Tap the box where each problem belongs.
Examples are the spec
Your training examples are the specification. The model copies what they do, mistakes included.
Cover easy, angry, confusing and off-topic messages. Check every example: sloppy examples teach sloppy replies.
Spot the bad example
The refund window is 14 days. Tap what this training reply would teach wrongly.
Hold some back
Before training, set some examples aside. The model never sees them while it trains.
They’re your test. Doing well on them means it learned the style, not just the examples.
How long to train
Train too little and nothing changes. Train too long and it memorises.
Check what it forgot
Then rerun your evals, including ones about skills you didn’t train.
Training hard on one task can make a model worse at others. Compare before and after.
Careful data, held back
Fine-tune for consistency: careful examples, a held-back test, and evals before and after.
Memorising instead of learning is called overfitting. The held-back examples are the validation set.