Modal describes it as a package of training building blocks, tuned defaults, and observability tooling; the company says it is available now.
The library accepts a model, dataset, and training recipe. Modal says datasets can come from Hugging Face, Harbor environments, or a team's harness and evaluation data. Recipes specify hardware, parallelism, rollout volume, environments, and other training parameters. Modal says its recipes have been validated from 0.5-billion-parameter models to multi-trillion-parameter models, with support for user-supplied models.
Dojo includes a dashboard for loss and reward curves, framework metrics, rollout trajectories, substep profiling, and container logs. Its CLI can return a run's status, metadata, logs, and agent traces. The source does not provide independent benchmarks, pricing, or reproducible evidence for its claims about training performance or customer results.