Overview

Fione is an open-source AutoML server. You upload a CSV file in the browser, Fione trains a model on it, and you use the model to make predictions or to serve it as a prediction API. It runs with one docker compose up.

You do not need machine-learning experience to get a first model. The EasyML wizard asks what you want to predict and how long Fione may work, then reports the result as a 1 to 5 star rating with a plain-language explanation. When you want more control, the detailed view opens every statistic, training setting and diagnostic.

Who Fione is for

• People new to machine learning who have a table of data (a CSV with a header row) and want a model and its predictions without writing code.
• Data scientists and analysts who want to train models quickly, read the cross-validation diagnostics and the model explanations, and script everything from Python when they prefer.
• Teams who want one shared place for datasets and models, with administrators, editors and viewers, and a prediction API they host themselves.

Features at a glance

• EasyML wizard. Upload, data analysis (the column to predict, the input columns, the prediction type, the maximum run time), run, prediction summary, predictions as a CSV. See EasyML wizard.
• AutoML screens. Column statistics and charts, correlations, leak warnings, every training setting, the job list, leaderboards, variable importance, score history, CV metrics, and model export and import. See AutoML screens.
• Model explanations. Diagnostics from the cross-validation predictions, importance and partial dependence curves, the contributions behind each prediction, and warnings for columns that predict the answer suspiciously well. See Explanations and leak warnings.
• Its own AutoML engine. Gradient boosting (LightGBM, XGBoost, CatBoost), random forest and extremely randomized trees, a linear model, a neural network and, for data with text columns, a TF-IDF linear model. Models are tuned with Optuna and ensembled within the run time you set, and Fione builds derived features and keeps only those that clearly help.
• Fione Serving. A small prediction API (/invocations, /model, /ping) for any trained model, run from one shared Docker image and the model files. See Fione Serving.
• Users and sharing. Administrators, editors and viewers; projects are shared per user. API tokens and a REST API (/api/v1) for scripts. See Users and sharing and REST API.
• Python modules. Data operations such as built-in frame operations, k-means and low-rank factorization; administrators can upload their own. See Python modules.
• Python client. fione-sdk drives Fione from your own scripts and notebooks. See Python client (fione-sdk).
• Assistant. Answers questions about a project's datasets and models and suggests a training run that a person confirms. See Assistant.
• Storage. Datasets and models live on SeaweedFS or any S3-compatible service; metadata is kept in SQLite.
• English and Japanese, light and dark themes, usable on phones.

How the pieces fit

A Fione installation has a few moving parts, all started by Docker Compose:

• The Fione application (the fione container) serves the web app and the REST API and runs the job queue, as a single process. Training, predictions, explanations and exports are jobs: they run on the server, so you can close the page and come back later. By default one training runs at a time and the others wait in the queue.
• Storage. Datasets and models are stored in SeaweedFS (the seaweedfs container), which speaks the S3 protocol. The database, the secret key, job logs and temporary files live in the application's data volume.
• The assistant's language model is a separate server that you provide: Ollama or Claude through the Anthropic API. The assistant is optional, and the rest of Fione works without it.
• Fione Serving is a second image, built on request. It serves one downloaded model as a prediction API on a port of your choice, separate from the Fione application.

Browser / scripts ──► fione (web app, REST API, job runner) ──► SeaweedFS (S3): datasets, models
                              │                                  data volume: SQLite, secret key
                              └──► language model (optional): Ollama or Claude

serving bundle (zip) ──► fione-serving ──► POST /invocations

Where to go next

1. Installation: start Fione, create the administrator and get a sample CSV.
2. EasyML wizard: train a model, make predictions and serve the model.
3. AutoML screens and Explanations and leak warnings: the detailed view.
4. Users and sharing, Python modules, Python client (fione-sdk) and Assistant.
5. For operators: Configuration, Operations, Fione Serving and Known limitations.