Siliqun RotorCrucible
Training data from your own data.
Crucible turns your documents into typed questions, answers and labelled customer messages, and your API or tool list into requests labelled call, ask or refuse for a model that chooses tools. A label is kept only when two independent reads agree, and every rejection is recorded with its reason.
Free on your own engine. Hosted generation on our models: contact sales.
How it works
Three steps.
Prepare
Bring your documents, or your OpenAPI file or tool list.
Shape and judge
Each label is read twice; a person settles the disputes and measures the rest.
Export
Export what passed, for your trainer or for fine-tuning, with the review report.
Checked twice
A label is kept only when two reads agree.
Every row is judged by two independent reads. When they disagree, the row goes to review with the reason, so what you export is what both agreed on.
Features
What you get
Two recipes today
From documents: questions, answers and labelled messages. From your tools: call, ask and refuse requests.
Two reads per label
A label is kept only when two independent reads agree.
Coverage before export
Every tool must have requests to carry out, to ask about and to refuse; a dataset that is short is not exported.
An honest figure
A review queue and an accuracy figure with its confidence range.
Your trainer's format
Our tool head's sets, chat fine-tune files in OpenAI and Anthropic form, and JSON Lines.
Stays with you
Your data and rows stay in your browser and on your engine.
Plans
Crucible on each plan
Decisions on your own engine are never counted. See pricing
Questions
Before you start
Where does my data go?
To your browser and your own engine. Hosted generation on our models is available on request.
What can it read today?
Documents (HTML, Markdown, plain text) and tool schemas (OpenAPI in JSON or YAML, MCP tool lists, OpenAI and Anthropic tool declarations).
What does it export?
From documents: typed questions, answers and labelled messages. From tools: requests labelled call, ask or refuse, as tool sets or chat fine-tune files, each with its review report.
Training data from your own data.
Training data from your documents and your own tools, checked twice.