Describe and design · Autessa Synthetic Data
LLM training data and prompt pairs
Autessa Synthetic Data generates LLM training data. The "Training pairs preset" sets up system, prompt and response columns, and the job exports the pairs as a JSONL or CSV dataset for fine-tuning, testing or evaluating an assistant.
The problem
Good training and test conversations are hard to collect
Teams building assistants need hundreds of realistic user prompts with ideal answers, including the rude, vague and off-topic ones. Real transcripts carry personal data and rarely cover the edge cases, and writing them by hand does not scale.
What you get
- One click sets up three columns: system (the instruction that frames the assistant), prompt and response.
- Prompt and response are required, so every pair is complete.
- Export as JSONL, one pair per line, or CSV.
- Up to 5,000 pairs per job, with appends for more.
- Describe the persona, topics and tricky cases; add sample pairs to set the voice.
- Near-duplicate prompts are removed across the whole job.
How it works
How does lLM training data and prompt pairs work?
Choose Dataset file
Pick the "Dataset file" target and the JSONL format.
Apply the preset
Press "Training pairs preset". The columns are filled in and the name defaults to "Training pairs".
Describe the assistant
Write who the assistant is, what users ask, and which hard cases to include, such as refunds, angry users and out-of-scope questions.
Review and export
Review the staged pairs, exclude weak ones, approve, then press "Export dataset" to download the file.
Every option
The training pairs preset
| Option | What it does |
|---|---|
| system | Long text, optional. "System prompt / instruction that frames the assistant." |
| prompt | Long text, required. "The user prompt." |
| response | Long text, required. "The ideal assistant response." |
| Format | JSONL (default) or CSV. |
| Rows | 1 to 5,000 pairs per job. |
Permissions
Who can do what?
| Who | Can |
|---|---|
| People in your company | Create datasets, see the company list of jobs, open any job and review, approve, append, cancel or discard it. |
| API keys | Can run quick generate, propose schemas and run dataset file jobs. Table targets need a signed-in person. |
Where AI does the work
How the generator handles this
Each batch writes complete pairs that follow your description and examples. Pairs missing a prompt or response are repaired or dropped, and repeated prompts are removed.
Generation runs on your company default model, and every row is checked against your columns before it is staged or written.
Your model and PII stripping →Connected
How lLM training data and prompt pairs connects to the rest of Autessa
FAQ
Questions about lLM training data and prompt pairs
What format is the JSONL file?
One JSON object per line with system, prompt and response keys, the format most fine-tuning and evaluation tools accept. Choose CSV if you prefer a spreadsheet.
Can I add my own columns to the preset?
Yes. The preset only fills in the three columns. Add more, such as a topic or difficulty label, in the schema editor.
Is it only for fine-tuning?
No. Teams also use the pairs as test prompts for assistants, examples for prompt design, and seed data for evaluation sets.
100% customizable, code and all
Make Autessa Synthetic Data fully your own.
Use it out of the box, then change anything. Ask Autobot, Autessa’s AI builder, for a new field, screen, report, workflow or agent and it builds it into your copy of the app. Or open the code and edit it yourself. Nothing is locked: the app, its data and its code are yours.
Start free and make it yours