Configure Data Collection

Start from a working first run. Adjust collection settings in config.yaml, then launch a new run and inspect its graph in the viewer.

Explore more candidates

Replace the generate block with:

generate:
  seed: 42
  nb_branch: 2
  nb_env: 1
  nb_max_update: 2
  max_start_per_stage: 5
  randomized: false

This requests two candidates per sampled situation and allows more trajectory starts per stage. Keep one environment until the task and service connections work reliably. Use a new run name:

magma-gen run branching_buttons --preset press_button.ButtonPressOrdered \
  --config-path ./config.yaml --no-coaching --no-judge --no-randomized \
  --sequence_of_buttons sw0,sw1 --instruction "Press sw0 then sw1."

Select branching_buttons in the viewer. Compare the candidates' decisions and stage outcomes. Multiple candidates can contain the same answer; increasing nb_branch does not guarantee different choices.

Adjust concurrency and exploration

Configuration	CLI override	Use it to…
generate.nb_branch	--nb-branch 2	Request more candidate decisions per situation
generate.nb_env	--nb-env 4	Run more physical executions concurrently
generate.nb_max_update	--nb-max-update 4	Increase the manager's simultaneous node-update capacity; minimum 2
generate.max_start_per_stage	--max-launch-per-stage 5	Bound trajectory starts per stage
generate.seed	--seed 42	Seed task creation and vocabulary randomization
generate.randomized	--no-randomized	Disable public vocabulary randomization

Increase concurrency gradually while watching GPU memory and agent/planner response times. More environments do not change how many candidates are requested from one situation. Increase the stage launch limit when alternatives stop before reaching later stages. Scenario-defined tool-call limits still apply.

For a slow agent, add --agent-timeout 600 or set magma_agent_timeout: 600 at the top level of the configuration. The timeout covers discovery and inference.

Collect procedural task variations

Use --definition once you have a tested task definition:

magma-gen run button_variations --definition my_buttons.Definition \
  --config-path ./config.yaml --max-total-target-tool-calls 10 \
  --no-coaching --no-judge --no-randomized

This command requires the installed definition from Build Tasks with Requests. Choose either --preset or --definition for each run. The old --scenarios option is unavailable.

--max-total-target-tool-calls controls the sum of declared stage targets used to construct one task (default 40). Requests are kept whole; the first request can exceed that budget. Each sampled request must produce a positive total target, even if some of its stages have a zero target.

Enable scenario skills or coaching

Pass registered skill keys with --skills, for example --skills PressPair for the two-button skill. Install the provider and register the skill before running the command.

To collect corrected alternatives, configure coaching. For interventions you write yourself, switch the viewer to Coach mode.

Inspect and export

Check the stages reached and the outcomes of the recorded candidates before exporting the run. A finished generation can contain failed or unfinished branches and can produce an empty dataset.

Read collection and branching for the explanation of branch state, simulation slots, and outcome classifications.