Skip to content

Roadmap

CuratorKIT 1.0 is the first public release. The pipeline architecture (readers, normalizers, generators, gates, and exporters behind stable base classes) is designed so each area below extends without breaking existing configs.

Directions under active consideration, in no particular order:

Sources and ingestion

  • Additional connectors: web pages, object storage, SQL databases
  • Incremental ingestion for growing corpora

Generation and gates

  • Additional generation tasks and multi-document grounding
  • Configurable gate ensembles and per-domain judge rubrics
  • Cost tracking and budget caps per run

Outputs and integrations

  • Additional export formats as trainer ecosystems evolve
  • Direct dataset publishing to the HuggingFace Hub with generated dataset cards
  • Deeper AlignTune integration beyond the documented curate-then-train workflow, such as one-call handoff without a Hub round-trip

Priorities are driven by usage. Propose or upvote items in GitHub Discussions; concrete proposals are welcome as feature requests. No dates are committed here; the changelog records what ships.