SIDERIUSSIDERIUS

Get started

Install, check, then ask your agent

About ten minutes. Nothing spends API credit or GPU time until step 5.

  1. Step 1: Install

    git clone https://github.com/yuema137/SIDERIUS.git
cd SIDERIUS
uv sync --group dev --frozen

    Expected result: a .venv/ inside the checkout, pinned by uv.lock.

  2. Step 2: Check it, with no key, data or GPU

    .venv/bin/python -m pytest \
  tests/unit/examples/test_quickstart_pack.py::test_composition_authority_resolves_from_this_checkout \
  tests/unit/examples/test_quickstart_pack.py::test_shipped_manifest_composes_with_the_declared_values -q

    Expected result: two passing tests.

  3. Step 3: Ask your coding agent

    Pick a style and paste it. Replace the example paths with yours.

    Professional: you lead the plan

    Use SIDERIUS in Professional style. My MNIST images and labels are in /data/mnist.
Build a digit-classification task and a three-iteration fixed-workflow experiment
in ~/mnist-research. Keep the test split out of search feedback; optimize validation
accuracy. Help me set model routing and GPU/time budgets, then show me the saved
configuration, launch script and result paths before running.

    Assistant: your agent guides you

    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.
    Tutorial: configure your task

    Expected result: your agent asks about splits, metric and budgets, then saves a task package in your project.

  4. Step 4: Preview, without running

    bash scripts/launch/run_chain.sh \
    --mode lilab \
    --workspace /path/to/your/workspace \
    --run_name first_run_v1 \
    --task_composition configs/task_composition/quickstart.yaml \
    --data_dir /path/to/workspace/quickstart_data \
    --num_iterations 1 \
    --max_rounds 1 \
    --dry-run

    Expected result: the exact child commands, with no side effects.

  5. Step 5: Launch when you decide

    Real runs make paid calls. Bind only the keys your routing needs, in the launching shell.

    First-run guide and required flags (opens in a new tab)

    Expected result: a workspace with records, Health config and generated models.

Prefer a ready-made task? Start with the TESS notebook tutorial in siderius-exp.

TESS tutorial (opens in a new tab)

Source of truth: SIDERIUS getting-started guide (opens in a new tab)

Next: how it works · tutorials · API keys (opens in a new tab)