SIDERIUSSIDERIUS

Tutorial 2 of 3

Configure your own task with your coding agent

Describe your data and goal; your agent saves a task package in your own project.

The idea to keep in mind

You own the science; your agent owns the typing. Decisions such as the split, the metric and the validity checks change what a result means, so your agent asks you about them. Everything it writes goes into your project, never into the SIDERIUS checkout.

Before you start

  • The SIDERIUS checkout installed with uv sync --group dev --frozen (see Get started).
  • A coding agent that can open the checkout and read its files.
  • Your dataset location and a new, empty place for your project outside the checkout.

Source of truth: the repository's getting-started guide (opens in a new tab)

Steps

  1. Step 1: Open the checkout in your coding agent

    Your agent will follow the repository's setup guidance (opens in a new tab). You do not need to write configuration files first, or know what a manifest is.

    Expected result: your agent can read the repository. No global skill installation is needed; the setup guidance lives in the checkout.

  2. Step 2: Describe your problem in one or two sentences

    Say what you have, where it is and what you want to predict. For example, in Assistant style:

    Use SIDERIUS in Assistant style. I have MNIST images and labels in /data/mnist;
help me build and evaluate a digit classifier. Follow this repository's setup
guidance, help me choose workflow or orchestration, and prepare an external
project with the environment, task, experiment and launch script.

    Prefer to lead? Use Professional style and state your plan; the prompts are on Get started.

    Expected result: your agent inspects your data and hardware and explains anything missing, such as a dependency or a credential name.

  3. Step 3: Agree on the scientific choices

    Answer the questions that matter; your agent reuses choices you have already given. Take your time here. Changing the metric changes the task; giving the same task more search time only changes the experiment.

    Expected result: concrete proposals with reasons: inputs and target, a leakage-safe split, which data guides the search and which stays held out, the metric, and any task-specific validity rules.

  4. Step 4: Let your agent write the task package and experiment

    The task package (opens in a new tab) declares the science: data, model inputs and outputs, objective, metrics and validity checks. The experiment holds the run choices: model routing, workflow, iterations and budgets.

    /absolute/project/                 # outside the SIDERIUS checkout
  task.yaml                        # task manifest; references your task plugins
  task_plugins/                    # dataset/model/metric implementations
  data/                            # existing dataset, or configure another external root
  llm.json                         # saved provider/model profile
  scripts/                         # reviewed experiment launch scripts
  workspaces/                      # scientific run outputs

    Expected result: files in your project, for example the layout below. These names are illustrative.

  5. Step 5: Check offline and preview the command

    For a first process check, the guidance recommends a small, low-cost test profile and a small run: one iteration, bounded data. An optional configuration review (opens in a new tab) can produce an HTML report of the setup.

    Expected result: your agent reports the offline checks it actually ran, the saved task and experiment paths, the key settings, the exact launch command and where results will be written.

  6. Step 6: Decide when to run

    Launch from your project's saved script, or use the browser console (tutorial 1). Then follow it with tutorial 3.

    Expected result: nothing runs until you say so. A real run uses API credit and GPU time, within the limits you agreed.