Fine-tuning data is a set of example conversations: a prompt and the ideal response written by a person. The model learns to answer like those examples, in their accuracy, structure, tone and judgment. Because it copies the examples so directly, a small set of excellent examples is worth more than a large set of average ones.
Research backs this up. In the 2023 LIMA study, a large model fine-tuned on just 1.000 carefully chosen prompts and responses produced answers that people judged equal to or better than GPT-4's in 43% of cases (Zhou and others).
What Makes a Good Example
A good prompt is realistic: it sounds like a real user, with real context, ambiguity and imperfect wording. A set full of tidy textbook questions teaches the model to handle tidy textbook questions.
A good response is:
- correct, with every fact, number and step checked
- responsive, answering what the user actually asked, in the format they asked for
- complete but not padded, with no filler introductions or repeated summaries
- in the right voice, matching the product's tone guide
- honest about limits, saying what it doesn't know and asking for missing information when needed
A worked example.
Prompt: "hi, I was charged 15 euro for a transfer to my mum in Spain, is that normal?? I thought EU transfers were free"
A weak response:
Thank you for reaching out! I completely understand your concern. Transfer fees can vary depending on many factors. Please contact our support team for more information about fees. Is there anything else I can help you with?
Polite, but it answers nothing and sends the user away.
An ideal response:
A €15 fee on a euro transfer within the EU is higher than our fee for a standard SEPA transfer. Two things usually explain it: the transfer was sent as an urgent (instant or priority) payment, or it was made in another currency. You can check which in the transaction details in the app. If it was a standard SEPA transfer in euro, reply here with the date and amount and I'll pass it to the team to review the fee.
It answers the question directly, gives the likely reasons, tells the user how to check and offers a concrete next step.
Note: the example is illustrative; a real project's guidelines define which fees and next steps the assistant may mention.
Covering the Right Prompts
The model learns from what it sees, so the mix of prompts matters as much as the quality of each answer.
| Cover | Why | Example |
|---|---|---|
| The real spread of topics | The model should be good at what users actually ask | Pull topics from real support logs, anonymised |
| Easy and hard cases | Hard prompts teach judgment; easy ones teach format | A simple balance check vs. a disputed charge abroad |
| Ambiguous requests | The model should ask a clarifying question when needed | "Can you move my money?" (move where?) |
| Requests to refuse or redirect | The model must learn where its limits are | Asking for another customer's details |
| Multi-turn conversations | Real users follow up, change their mind, correct the assistant | A second message saying "no, I meant the other card" |
| Different user styles | Typos, slang, short and long messages, several languages | "wht r the fees 4 transfer" |
| Formats | Lists, tables, code, short answers, step-by-step instructions | "Give me the steps as a numbered list" |
Avoid near-duplicates. Fifty versions of the same question teach less than fifty different questions. Check the set for repetition before training.
The Writer's Style Guide and Review Process
Writers need a style guide, just as labelers need annotation guidelines. It should cover:
- Voice and tone: formal or friendly, first person or not, how to apologise, words to avoid
- Length: default length, when to be longer, when one sentence is enough
- Structure: when to use lists, headings, tables or code blocks
- Facts and sources: what may be stated, what must be checked, how to handle uncertainty
- Boundaries: what the assistant must refuse or redirect and how to say it politely
- House rules: product names, currencies, date formats, legal disclaimers
Review every example. Fine-tuning copies mistakes faithfully, so each example should be checked by a second person before it enters the set:
- A writer drafts the response.
- A reviewer checks it against the style guide and fact-checks it.
- Disagreements go to a lead and the decision is added to the style guide.
- A sample of approved examples is re-checked by the client.
Use experts where the content needs them. Medical, legal, financial and technical answers should be written or reviewed by people qualified in the field. See Expert Annotation.
Why Human-Written Means Human-Written
It is tempting to have another AI model draft the responses and ask people to tidy them up. For some projects that is a deliberate choice, but it changes what the data is. The new model then learns from another model's habits, mistakes and blind spots, not from human judgment. Writers who quietly paste AI drafts make the problem worse, because nobody knows it happened.
If a project needs human-written data, say so in the contract and the guidelines, block pasting in the writing tool and check for AI-written text. At AnnoTrue, client data is always written and labeled by people, never by AI.
Examples Written by People
AnnoTrue writers are tested on writing and fact-checking, follow your style guide and every example is reviewed by a second person. Try it with a free pilot.
Sources
- Zhou and others, LIMA: Less Is More for Alignment, 2023