Latest in AI
Modal Dojo Explained: LLM Training Recipes, Models and Evaluation
Modal's new training library packages the setup and observability around post-training, leaving teams with the harder question of what to reward.
Modal launched Dojo on October 21, a library for training language models on Modal with model-specific recipes, a Python interface and an observability dashboard. It reduces the configuration work around post-training. Teams still need a suitable dataset, a useful objective and a way to verify that training improved the model.
The announcement describes an available product built on open-source training frameworks, including Miles and Slime. Its starting point is a model, data and a recipe, rather than an infrastructure project that must be assembled before the first experiment.
What goes into a Dojo training run
The quickstart requires Python 3.12 and the modal-dojo package. Setup configures Modal and the dashboard; an optional skills bundle gives coding agents guidance for working with the library.
A TrainConfig joins a model, dataset and recipe. The documented example loads prompts and labels from a Hugging Face dataset, supplies a custom reward function and launches the run. Model support includes several Qwen families as well as DeepSeek, GLM, Gemma, Kimi and other open-weight options.
That list is a place to start selecting a model, not a ranking of which one will work best. A smaller model that already handles most of the task may be a more useful baseline than the largest model a training system can accommodate.
How recipes control training and cost
A Dojo recipe describes hardware, parallelism, training type, rollout size and the environment. The recipe guide exposes the underlying framework's settings instead of hiding them behind a fixed workflow.
It supports supervised fine-tuning and reinforcement learning. In supervised fine-tuning, the model learns from the dataset's conversations. In reinforcement learning, it generates responses that are scored by a reward function. The latter requires deciding what success means well enough to encode it.
The guide also makes a practical cost distinction: training and rollout generation can share a cluster, or inference can run on a separate cluster. Separating them creates another compute requirement. A short configuration file can still launch a substantial workload.
For an initial experiment, the useful question is whether the setup produces a measurable signal at modest cost. Increasing steps, prompts per step and samples per prompt together increases generated work multiplicatively. More experiments are only helpful if the results can be compared meaningfully.
The dashboard should explain the run
Dojo's dashboard documentation describes reward curves, score distributions, rollout traces and a breakdown of time spent in generation, training and other steps. Worker logs are available during and after a run.
These views answer different questions. A reward curve shows whether the measured objective is improving; a trace helps explain what the model actually did. Step timing helps distinguish a slow learning process from an experiment spending much of its time moving weights or generating long responses.
The distinction matters because a model can become better at earning a reward without becoming more useful. Inspecting examples alongside the aggregate score should be part of the experiment, not an optional debugging step after a surprising result.
Dojo versus writing your own training loop
Dojo is the higher-level option for teams that want a starting recipe and supporting tools. Modal's separately announced Spindle targets developers who want to drive individual training operations through a Tinker-compatible API. They address different levels of control.
One current limitation is that the recipe guide does not support passing eval_dataset for supervised fine-tuning. Evaluation needs its own plan. Keep test cases outside the training process, compare against the original model and inspect failure cases before deciding that a checkpoint should replace anything in production.
The meaningful milestone is a better result on a task the model was not rewarded for memorizing. Dojo can shorten the path to running that experiment; it cannot choose the right success criterion on a team's behalf.
I'm Paras. I work at the intersection of AI, engineering and distribution.
If you're building an AI product, I help turn technical launches and product thinking into content developers and founders can understand and share.