Generate Data with MAGMA-GEN

MAGMA-GEN collects training data by running an agent on tasks in simulation. It records the agent's decisions and their outcomes in an interaction graph. Optional coaching proposes corrections that GEN executes and checks.

Your first run

Use a provided button task and full-history with Starter MAGMA to check the complete workflow:

1. Install MAGMA-GEN, the agent package, and the planner.
2. Configure your model and services.
3. Launch a first generation.
4. Inspect the run to see what the agent did and whether the task succeeded.
5. Export the saved graph and inspect the training examples.

The first run uses one environment and one branch, with coaching and semantic answer checks disabled. It checks the integration; task success depends on your model. A short or unsuccessful run may produce an empty dataset.

Model setup

Configure Starter MAGMA with its Hugging Face identifier; full-history downloads the weights on first launch. You can also use the provided remote agents or your own agent integration.

What you get

Task + agent + generation settings
  → agent decisions executed in simulation
  → saved interaction graph
  → selected examples rendered by the agent exporter
  → training and validation datasets

GEN records experience and exports data. Model training happens separately; a generation run does not update the agent's weights.

Understand and extend the workflow

After your first run, configure collection to explore more branches, or enable coaching. The viewer lets you submit human corrections while a run is active.

For the explanations behind these operations, read collection and branching and coaching and corrections.

To collect data for your own application, create a scenario or connect your model or agent. You can change either independently.

Software and paper versions

These guides cover the stable v2 packages. The software version used for the MAGMA-GEN paper is v0.1; its commands and implementation details may differ from this version.