machinable is a content-addressed system for research code. You write an experiment once; running it, reloading its results, and asking questions of them are the same gesture. A run's identity is its module and configuration, so identical runs deduplicate, sweeps are incremental, and results stay reachable from code with no run names or paths to track.

from machinable import get

get("train", {"lr": 0.1}).launch()   # runs it ...
get("train", {"lr": 0.1}).loss()    # ... or reads it back

Installation

pip install machinable

Documentation

Visit machinable.org to get started.