Ludwig

Ludwig starts from an unusual premise for a deep-learning tool: that most models are assembled from the same handful of parts, and what really changes between projects is the shape of the data. So you describe the data, not the network.

Declarative models

A Ludwig experiment is a YAML file naming input features and output features with their types — text, category, number, binary, image, audio, sequence, date. Each input type has a default encoder, each output a default decoder and loss, and Ludwig wires them together through a combiner in the middle. The entire model below is four lines of configuration, trained with ludwig train:

input_features:
  - name: review
    type: text
output_features:
  - name: sentiment
    type: category

Where the AutoML part comes in

Every default is overridable in the same file — swap the text encoder for a pretrained transformer, set a learning-rate schedule, add a second output head — so the ceiling sits well above the four-line entry point. Hyperparameter search is declared in that config too and executed through Ray Tune, while ludwig experiment runs the train-evaluate-compare cycle and auto_train picks a whole configuration from the dataset for you. Originally built at Uber and now maintained in the Linux Foundation AI ecosystem, Ludwig earns its place where the task is multimodal or multi-output — classifying free text alongside tabular columns, say — a case where general tabular AutoML libraries have little to offer.

Ludwig is an open-source declarative deep-learning framework: you specify input and output features with their data types in a YAML file, and it builds, trains and evaluates the model. Per-type defaults can be overridden in the same file, and hyperparameter search runs through Ray Tune. It matters as the low-code route to multimodal and multi-output models — text, images, numbers and categories in one network — without writing model code.

Ordering a sandwich is a matter of saying what goes in it, not explaining how to butter the bread. Ludwig brings that style to artificial intelligence.

Building a neural network — the kind of program behind photo and language AI — normally means writing a great deal of code. With Ludwig you write a short list instead: here is what I am giving the computer (some text, a picture, a few numbers), and here is what I want back (a category, a score). It assembles a suitable network and trains it for you. People reach for it when their information is mixed together, like product reviews alongside prices and photographs.