Token Factory

Fine-tune open models, manage training datasets, and serve inference endpoints on Saturn Cloud

Token Factory is the Saturn Cloud product surface for the open-model lifecycle: upload a training dataset, fine-tune an open base model, and serve the result as an inference endpoint. It is built on the same compute, storage, and scheduling primitives that run the rest of the platform, exposed through a higher-level API that speaks in datasets, jobs, and endpoints rather than pods and volumes.

When to use Token Factory

Use Token Factory when you want to fine-tune and serve open-weight models without assembling the training stack yourself. You provide a dataset and a base model; the platform renders the training configuration, schedules the GPU job, captures the resulting checkpoint, and (for serving) stands up an inference endpoint that reads that checkpoint.

If you instead want full control over a custom training loop, run a Job directly and manage the environment yourself. Token Factory trades that flexibility for a managed, repeatable path through the most common fine-tuning and serving workflow.

The three surfaces

SurfaceWhat it isPage
DatasetsTyped, validated training data registered with the platform and shared across your organizationDatasets
Fine-tuning jobsA GPU training run over a base model and a dataset, producing a checkpointFine-Tuning Jobs
Inference endpointsA persistent deployment that serves a checkpoint over an HTTP APIInference Endpoints

These connect in a line: a dataset feeds a fine-tuning job, the job produces a checkpoint, and the checkpoint backs an inference endpoint. The Concepts page explains the objects and how they relate.

Organization-level by design

Token Factory objects are owned by the organization, not by an individual user. A dataset you register is visible to everyone in your organization; a checkpoint a teammate produces is one you can serve. This matches how teams actually work with training data and models: the data and the resulting weights are team assets, and the platform treats them that way.

Getting started

  1. Register a dataset of training examples.
  2. Start a fine-tuning job against a supported base model.
  3. Serve the checkpoint as an inference endpoint.

The pages in this section document the API surface and the behavior you can rely on at each step.