FACTS → QUESTIONS → REASONING

Questions with evidence.

Build LLM training examples whose answers you can trace.

Conceptuum knowledge graph

Review & export

Preparing the first batch…

TEACH REASONING FROM FACTS

The evidence travels with the answer.

Every training example contains instructions, a question with graph facts, and an answer. Concept and edge IDs preserve the exact meaning.

Training JSONL contains a messages array with system, user and assistant roles.

What gets checked?

Source fidelity: Every supplied edge comes from the selected graph context.
Independent recomputation: A separate solver derives the answer from the input facts. Algebra checks catalog and property answers against the full graph.
Knowledge boundaries: Unknown is different from negative. Missing catalog members do not prove real-world incompatibility.

Generator & format guide ↗

Review before training.

Automatic checks establish consistency with supplied facts. They do not certify every assertion about the world. Known problematic source assertions are filtered.

Uncheck examples you do not want to export. This selection stays on this page; download it before leaving.

Keep evaluation separate.

Keep the same group_id in one split. Inspect shared concepts and evidence across splits as well; groups alone do not prevent all leakage.

Generate larger batches locally with the Python CLI.

Questions with evidence.

Generate a dataset

Review & export