When the default model isn’t enough
Most of the time, the AI you use out of the box is great. It answers, summarizes, drafts, plans, and reasons well.
But sometimes “great” isn’t quite right. You want answers that sound like your team. You want summaries that use your jargon. You want a model that knows your products, your customers, your processes — not in the abstract, but in the same words your team uses every day.
That’s the moment to fine-tune. And the first step in fine-tuning is the most important: clean, exportable training data, built from what your team actually does.
You don't have to be a machine-learning engineer to prepare a dataset. The hard part — clean data, consistent structure, real examples — is something your team is already producing in the course of normal work. This page is about capturing that into a form the training step can use.
Who this is for
- Product, customer success, and operations leads who want an AI that speaks in their team’s voice.
- Data leads preparing training data for a fine-tuning project.
- Workspace admins building reusable training sets that the rest of the team can draw from.
You won’t write code. You will make a few choices about structure and source.
What a fine-tuning dataset is
A fine-tuning dataset is a list of example pairs. Each pair has a prompt (the kind of question or instruction the AI will be asked) and a completion (the kind of answer your team would give).
A few examples that make it concrete:
| Prompt | Completion |
|---|---|
| ”Draft a renewal note for an enterprise customer at the 60-day mark.” | A renewal note in your team’s voice, with your team’s structure. |
| ”Summarize this support ticket for the on-call engineer.” | A summary in the format your engineers actually use. |
| ”Write a one-paragraph product update for the weekly digest.” | A paragraph that reads like your team wrote it. |
Done well, a fine-tuning dataset captures the tacit standards your team operates on — the things you’d struggle to put in a style guide but everyone recognizes when they see it.
How VDF AI Data helps you build one
Building a great dataset by hand is slow. VDF AI Data turns the slowest parts into a few clicks.
Start from a connected source
Pick a table, a document folder, or a feature list. Your dataset draws from real, current data — not a static export.
Pick a mapping template
Choose how rows turn into prompt/completion pairs. Common templates ship ready to use; you can also define your own.
Preview before you build
See the first few generated pairs before you commit. Catch a mapping mistake in a minute, not a thousand rows later.
Track dataset status
Draft → Ready → Exported. Always know where each dataset is in its lifecycle.
Export in standard formats
JSONL for most training workflows. CSV for inspection and editing. Both straight out of the screen.
Build many datasets per source
The same asset can back several datasets — one per use case, one per audience, one per output type.
Mapping templates, in plain language
A mapping template is how you tell VDF AI Data to turn a row into a training example.
The simplest template: “this column is the prompt, that column is the completion.” For an FAQ asset, that might be one column called question and another called answer.
More useful templates combine columns:
- Header + body becomes the prompt; resolution becomes the completion. Great for support tickets.
- Document title + summary becomes the prompt; full document becomes the completion. Great for a team that wants the AI to expand outlines into full drafts.
- Customer attributes become the prompt; account-manager notes become the completion. Great for tuning an assistant on how your team writes about customers.
The screen ships with templates for the most common shapes. You can adapt one, or define a fresh one for an unusual source.
The mapping matters more than the volume. A small, well-mapped dataset (a few hundred sharp examples) outperforms a giant, sloppily-mapped one (tens of thousands of noisy examples). Get the template right; build small first; expand once you trust the shape.
Building a dataset — the workflow
-
Pick your source.
A connected database asset, a document folder, or a feature list. Start with something where the columns or content are clearly "prompt-like" and "answer-like."
-
Choose a mapping template.
Pick one of the prebuilt templates or define your own. Match the source's natural shape.
-
Preview a sample.
Look at the first few generated pairs. Are they clean? Are they representative? Adjust if not.
-
Build the dataset.
VDF AI Data reads the source, applies the template, and assembles the pairs. Status moves from Draft to Ready.
-
Spot-check the result.
Open the dataset, scroll, search. Look for missing fields, mangled formatting, or obvious noise.
-
Export.
JSONL for most training workflows, CSV for review. The dataset moves to Exported state and is tracked in your dataset history.
Status, in plain language
| State | What it means | What to do |
|---|---|---|
| Draft | You’re still configuring source and mapping | Finish the form, run a preview, build |
| Ready | The dataset has been assembled and is reviewable | Spot-check, then export |
| Exported | A file has been produced; the dataset is captured | Use the export downstream; the dataset stays in your history |
How much data do you actually need
The honest answer: less than people expect. A focused dataset of a few hundred high-quality, representative examples often produces a tangible shift in how the model writes about your business. Tens of thousands of mediocre examples can produce less shift — because they wash out the signal.
A useful sequence:
- First dataset: 100–300 examples. Test the workflow end-to-end. Confirm the model picks up your team’s voice in the small.
- Second dataset: 500–2,000 examples. Once the small one works, scale up on the parts that worked.
- Ongoing maintenance: append, don’t replace. As new data lands, add to the dataset rather than rebuilding from scratch.
What good fine-tuning data looks like
Three quick checks before exporting:
- Consistent format. Every prompt looks like every other prompt. Every completion looks like every other completion.
- Real examples. No synthetic placeholders. No “TBD” or “[customer name]” in production rows.
- Representative coverage. Easy cases, hard cases, edge cases. If your dataset is 90% easy questions, the tuned model will be great at easy questions and surprisingly bad at the rest.
Sensitive content needs deliberate handling. If your dataset includes customer-identifying information, contracts, or anything regulated, scope the dataset accordingly and review it with whoever owns data governance. Once exported, the file is just a file.
What happens after you export
The export is a starting point, not an endpoint. The exported file flows into a training workflow — which may live inside your own ML platform, or be handed to a partner who runs the training step.
VDF AI Data keeps the dataset record even after export: the source, the template, the build settings, the row count. That record means you can rebuild the same dataset in a month, with a slightly different mapping, against a slightly different scope — and compare outputs cleanly.
When fine-tuning is worth it (and when it isn’t)
A useful self-check before investing.
| Signal | Fine-tuning is probably the right tool | Probably not |
|---|---|---|
| The default model gets the substance right but the voice wrong | ✅ | — |
| Your team has thousands of consistent examples sitting in a system already | ✅ | — |
| You want the model to know your specific business facts | — | Use vector indexes instead |
| You want one-off answers, not a behavioral shift | — | Use Agents with a strong brief |
Fine-tuning shifts behavior. For lookup, search and grounding tools are usually better — they keep facts current.
Where to go next
- Vector indexes and semantic search — the other path: searchable, current, no training step.
- Features and relationships — scope a dataset to a focused feature list.
- Discovering your data — find candidate sources for training pairs.
- VDF AI Agents — sometimes the right answer is a sharper agent, not a tuned model.