Start with your material
ShuJian Cube is our open-source dataset generation platform. Its starting points include documents, agent context and open task briefs.
Before collecting examples, write down the problem you want the data to address. A concrete objective makes it easier to decide what belongs in the collection and what needs a separate task.
Choose the learning goal
The platform organizes tasks around CPT, SFT and DPO: learning from domain material, learning clear responses from examples, and learning from answer preferences.
When preparing a brief, describe the knowledge, the intended behavior and the selection criteria separately. This helps reviewers assess each example against a clear purpose.
Make room for review
The documented workflow moves through source, planning, generation, review and sharing. Teams can follow progress and inspect examples before sharing a collection.
Fluent writing still deserves scrutiny. Check the source, the surrounding context and any conditions that an answer may have omitted.
Work in the open
Explore the repository to see the current workflow and project documentation, then begin with a small collection suited to your own task.
Explore ShuJian Cube on GitHub ↗
Product details in this note are based on the public README and workspace interface.
