Working with the suite

miForge is standalone. It loads or creates a model itself, trains in its own process and writes its own checkpoints. Every sibling application is optional, detected at runtime, and reported by name when it is missing. GET /api/health lists each integration and its state.

Application	What miForge uses it for	Configure with	Without it
miDataworks	Dataset parts for packs, under a versioned contract; the human-review queue for the human-review selector.	MIDATAWORKS_BASE_URL	Import parts from your own files; use rule, judge, reference or composite selectors.
miLLM	Serving miForge's checkpoints; scoring through its OpenAI-compatible endpoint (as one provider among many); the optional hot-update mode.	MILLM_BASE_URL, and a PROVIDER_* entry to score through it	Score in-process; serve checkpoints with any compatible server.
miStudio	Interpretability of the models miForge produces.		Nothing in miForge depends on it.

miDataworks: parts

miDataworks is where dataset parts are written, reviewed and versioned. miForge consumes its dataset versions under the midataworks.dataset-version/v1 contract (vendored, byte for byte, in miForge's docs/schemas/), and refuses anything that does not conform, naming the field.

• GET /api/v1/midataworks/status says whether miDataworks is configured and reachable, and what is unavailable without it.
• GET /api/v1/midataworks/available-parts lists the versions it offers, by kind.
• POST /api/v1/parts/import with "source": "http" imports one.

A part miDataworks did not publish is recorded as external and is never presented as miDataworks-sourced.

miLLM: serving and scoring

A miForge checkpoint is a standard model directory, so miLLM loads it like any other model. Nothing needs to be pushed into miLLM: point it at the checkpoint directory.

To score through miLLM, add it as a provider (see Scoring). miLLM is one OpenAI-compatible endpoint beside vLLM, Ollama and others; nothing in miForge treats it specially there.

Contracts

The formats miForge produces and consumes are published as JSON Schema files in the repository's docs/schemas/. See Artifacts and schemas.