Synthetic data
Generate the training and eval data other models won't write
Records in your own schema, including the hard negatives and harmful-looking examples a classifier must learn from. Unrestricted models write them in JSON mode, within the Acceptable Use Policy.
Unrestricted means fewer refusals, not no rules. Read the Acceptable Use Policy
Your schema, one record per reply.
Describe the record, plan the coverage and generate from your own script.
request.jsonJSON modeThe schema in the request
Turn on JSON mode and describe the record in the system prompt. Each reply is one object you validate.
record.jsonOne recordA hard negative
A message that looks like a scam but is legitimate: the kind a filter gets wrong.
coverage.csvCoverage planA coverage plan
List scenario by difficulty, mark what you hold out and generate for the thin cells.
How a data team uses it.
The pipeline is yours. The model fills the records.
You
Define the record and the coverage
Fields, label set, what makes an example hard, and which scenarios to hold out.
Modellane
Fills records in batches
One record per reply with JSON mode on, varying scenario, style and difficulty per call.
You
Filter before you train
Validate every record, remove duplicates and read a sample by hand. Keep held-out scenarios out of training.
What we do, and what we don't.
Modellane answers requests. Planning, running and checking a dataset happen on your side.
Read the Acceptable Use Policy
What we do
- Write records in your schema with JSON mode
- Write hard negatives and harmful-looking examples for classifiers
- Bill per token, so you pay only for records you request
What we don't do
- Run a generation pipeline or schedule jobs for you
- Validate, deduplicate, host or export datasets
- Store your prompts or the model's replies
- Allow uses the Acceptable Use Policy prohibits, whatever the model answers
Which models fit: Models whose id contains -unrestricted write the harmful-looking and adversarial examples that typical assistants refuse to generate. Standard models keep their base model's behavior, behind the same API. See every model
FAQ
Synthetic data questions
- Do you generate the dataset for me?
- No. You call the API from your own script; each request returns one or more records.
- How reliable is the JSON?
- JSON mode returns a single valid object per reply. Still validate every record against your schema and drop the ones that fail.
- Are generated records stored?
- No. Prompts and replies are never stored, so the only copy is the one your script saves.
- Will it write harmful examples?
- Unrestricted models write the examples a classifier needs to see. The Acceptable Use Policy still applies, and some content is prohibited for any purpose.
Fill the cells your dataset is missing.
Create a key, define the record and generate a first batch.