Repolode
Submit
google-research

google-research/kauldron

A modular, scalable library for training machine learning models in JAX/Flax, built for fast research iteration.

View on GitHubWebsite
Stars
359
Forks
36
Open issues
83
Last commit
yesterday

What is it?

Kauldron is a Python library for training machine learning models, designed around research velocity and modularity. Its components are self-contained and can be used independently of the framework, so teams can pair it with their own data pipelines, Flax models, and optimizers.

The library acts as glue between datasets such as TFDS, Grain, and SeqIO, any Flax model, and any optimizer. Configurations can be customized and overridden, for example to sweep over model architectures or change parameters of inner layers. Example configs such as an MNIST autoencoder serve as starting points.

It is built for a Colab-first workflow for quick prototyping. Built-in features include plotting integration, a profiler, post-mortem debugging, and runtime shape checking.

Key features

  • 01Works with TFDS, Grain, SeqIO, or custom data pipelines.
  • 02Accepts any Flax model and any optimizer.
  • 03Configurations can be swept and overridden at any depth.
  • 04Components can be used independently outside the framework.
  • 05Includes a profiler and runtime shape checking.

Use cases

  • Prototyping new model architectures in Colab notebooks.
  • Running hyperparameter or architecture sweeps for ML research.
  • Training Flax models on custom datasets with a standardized setup.

Things to know

  • The README states it is not an officially supported Google product.
  • The README gives no installation or run instructions.

Topics

#deep-learning#jax#flax#machine-learning#training-framework#python

FAQ

What is kauldron?
A modular, scalable library for training machine learning models in JAX/Flax, built for fast research iteration. Kauldron is a Python library for training machine learning models, designed around research velocity and modularity. Its components are self-contained and can be used independently of the framework, so teams can pair it with their own data pipelines, Flax models, and optimizers.
Is kauldron free to use? What license does it have?
kauldron is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is kauldron still maintained?
Actively maintained. The last commit to kauldron was on Nov 6, 2026.
What can I use kauldron for?
Prototyping new model architectures in Colab notebooks. Running hyperparameter or architecture sweeps for ML research. Training Flax models on custom datasets with a standardized setup.
Where can I get kauldron?
The source code and installation instructions are on GitHub: https://github.com/google-research/kauldron

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects