Modal ships Dojo, a post-training library that turns production agent traces into LLM training runs
Synopsis
Modal released Modal Dojo, a post-training library aimed at teams that already run agents in production and collect evaluations and traces. It bundles model selection, data loading, training recipes and metric observability behind one interface; the official example launches training of Qwen3_8_27B on 120 GSM8K training samples. Built on open frameworks such as Miles and Slime, it is said to be validated on models from half a billion to multi-trillion parameters.
Interpretation
Modal released a post-training library called Modal Dojo that lets teams train large models on data from their own production agents instead of building a training stack from scratch. The prior path was to rely on closed-source providers or rebuild training components; Modal argues that a deployed agent harness with evaluations and traces already covers most of the work, leaving only an entry point for training on proprietary data. The official minimal example uses the Qwen3_8_27B model, 120 GSM8K training samples from Hugging Face and a custom reward function, launching a run through TrainConfig and returning a training run ID.
Getting started takes under 100 lines of code: pick a model, load data, set a recipe, then launch. The model and dataset decide what is trained, while the recipe decides how training runs by specifying hardware, parallelism, total rollout size and environment. Modal says the library has been validated on models from half a billion to multi-trillion parameters and supports bringing your own; auto-optimized recipes are provided for all supported models.
The library ships an observability dashboard covering loss and reward curves, framework-emitted metrics, each rollout's agent trajectory, profiling down to the substep and container logs. These metrics are meant to catch reward hacking during training, debug stalled runs and track many concurrent runs. Modal also provides a CLI through which agents can read a run's status, metadata, logs and agent traces, backed by crafted skills that guide the agent.
Perspective
The report covers Modal's release notes for Modal Dojo, its example code and customer usage descriptions; the example launches training on 120 GSM8K training samples and does not report post-training model performance.
Modal says the library has been validated on models from half a billion to multi-trillion parameters but does not state the validation conditions; the customer cases describe usage without checkable performance numbers.