AutoML Automated Machine Learning

Problem

Tabular model tuning often depends on manual trial and error: learning_rate, num_leaves, cleaning, model switching, and ensemble strategies all require repeated runs and comparisons. This skill wraps CSV modeling into an autonomous experiment loop, letting an agent iterate on CPU toward a target metric and keep effective configurations.

How It Works

The flow starts from parameters: data_path, target_col, task_type, metric, time_budget, and max_experiments. It then checks the data, initializes a workspace, and sets a baseline. In the loop, it reads results.tsv, designs experiments by priority: hyperparameter tuning > model switching > feature engineering > ensembling; edits train.py, runs training, parses validation metric, memory, and runtime, then decides whether to commit or revert. It supports classification, regression, and multiclass tasks, with metrics such as accuracy, f1_macro, rmse, mae, and r2, and reports feature importance and overfitting diagnostics when available.

Boundaries

It is best suited to medium-scale structured data, limited compute budgets, and interpretable classical ML configurations. If the data has severe missing values, the target column is unclear, or the task requires deep learning or long-running model training, clean or curate the data first or use a different pipeline. The experiment loop runs autonomously until max_experiments is reached or the user stops it.