gmboquet/mixle

Mixle is a Python library for building statistical models out of mixed and messy data, things like numbers, text, categories, missing values, rankings, and even graphs. Instead of picking a specific model type yourself, you hand it your data and it works out what kind of model fits, then trains it. You can also hand it a specific structure to fit if you already know what you want. The library treats very different kinds of models, such as a classic statistical distribution, a neural network, or a model with hidden states, as the same basic kind of building block. Because of that, you can combine them and fit the whole combination with a single function call named optimize. For example, you could build a model with hidden states where each state is described by a different kind of smaller model nested inside it, and one call trains every layer together. One practical use the README highlights is distillation: pointing mixle at something slow and costly to run, like a large language model or an external API, so it learns from examples of that system's answers. The result is a small local model that handles easy cases on its own and only calls the expensive original when it is unsure, which can cut costs since the expensive system only gets used for a fraction of requests. Mixle can also wrap an existing PyTorch module and train it without you writing a training loop, batching code, or convergence checks yourself. The same model definition can run on plain NumPy, Numba, GPU through Torch, or a symbolic math engine, and can scale out to distributed systems like Spark, Dask, Ray, or MPI by changing a single argument rather than rewriting the model. Installing the base package via pip gives every distribution and local fitting. Extra features like GPU support, distributed backends, and specific data source connectors are added as optional installs. The project requires Python 3.10 or newer and is tested on Linux, with macOS working in day-to-day use though not formally tested in its automated checks.