Skip to content

Quick Start

Get Agentomics-ML running in a few minutes with the pre-built Docker image.

Prerequisites

Steps

1. Install the CLI and an Example Dataset

Install the package from PyPI, then download a single example dataset to try (AGO2_CLASH_Hejret2023). No repository clone is required:

python3 -m pip install agentomics
agentomics-download-dataset --dataset AGO2_CLASH_Hejret2023

2. Set a Provider Key

Export at least one provider key (or put it in a .env file in the current directory):

export OPENROUTER_API_KEY=...   # or OPENAI_API_KEY / ANTHROPIC_API_KEY

3. Run the Agent

agentomics-run --dataset AGO2_CLASH_Hejret2023

agentomics-run launches the Agentomics Docker image for you, reads datasets from ./datasets, and writes this run's results to ./outputs/<agent_id>/.

Drop --dataset AGO2_CLASH_Hejret2023 to pick a model, dataset, and iteration count interactively instead.

The validation metric defaults to AUROC for classification and MAE for regression. To choose one explicitly, pass --val-metric; list options with --list-metrics in place of the run arguments.

Using Your Own Dataset

Place your data in datasets/<your_dataset_name>/:

datasets/my_dataset/
├── train/
│   ├── input/          # Required: model input files
│   └── labels.csv      # Required: id,label
├── validation/         # Optional
│   ├── input/
│   └── labels.csv
├── test/               # Optional; hidden from the agent and evaluated afterward
│   ├── input/
│   └── labels.csv
└── dataset_description.md

The held-out test split is optional. It is not mounted into the agent worker. After a successful run, Agentomics starts a separate evaluation container with read-only access to the test split, evaluates the best iteration, and includes the results in reports. Run training with --dataset my_dataset. See Preparing Datasets for details.

Example Datasets

Download example dataset to try:

agentomics-download-dataset

List other available examples with:

agentomics-download-dataset --list

What Happens Next

The agent will:

  1. Prepare your dataset
  2. Run iterative ML development cycles
  3. Save the best model to the run's output directory (outputs/<agent_id>/)

Results include trained models, inference scripts, markdown reports in reports/markdown/, and PDF reports in reports/pdf/.

Next Steps