exege

Tools for understanding and evaluating scientific machine-learning models: what an emulator holds inside, sparse autoencoders trained on it, figures, and a local app over them. The name comes from the Greek stem exēgē-, associated with explanation and interpretation.

Research tool. exege is early. What is described here works; expect it to change.

Install

This is the light install: the code, and only Click. Anything heavier sits behind an extra named after the subpackage that needs it:

$ uv pip install 'exege-core[latents]'     # or: pip install 'exege-core[latents]'

For everything at once, install exege instead: it is this with every extra but nn (pip install exege, or 'exege[nn]' to add torch). Both provide the same import exege and the same exege command.

A missing extra says so, with the command that fits how exege was installed.

Try it, with no model and no data

A toy emulator writes a latent archive with nothing but numpy; everything else reads it like any other:

$ exege latents toy scratch/toy/control
$ exege latents info scratch/toy/control --mask-variable sst
$ exege latents region scratch/toy/control --lat 10 --lon -114 --time 2 --centered --pcs 2
$ exege latents fields scratch/toy/control --field precipitation --top 3
$ exege nn sae scratch/toy/control --features 64 --k 4 --out scratch/toy/sae.npz   # needs exege-core[nn]
$ exege app --latents scratch/toy                                                  # needs exege-core[app]

The CLI is a thin client of the Python API; anything it can do, a notebook can:

from exege.latents import Region, analyze_region, open_source

source = open_source("scratch/toy/control", mask_variable="sst")
result = analyze_region(
    source, time=2, layer=3, region=Region(lat=10, lon=-114, radius_km=1500), centered=True
)
result.ranking.channels  # the channels that respond most strongly there

More

• Documentation: https://mahf708.github.io/exege/
• Source and issues: https://github.com/mahf708/exege

BSD-3-Clause. exege.latents, exege.figures and exege.app grew out of the latent space visualiser for weather models (Tempest, Beylich & Craig 2026, arXiv:2604.20467, doi:10.1007/978-3-032-29915-4_10); see NOTICE. The sparse autoencoders follow MacMillan & Ouellette (2025, arXiv:2512.24440); the B-spline autoencoder of Cheon (2026, arXiv:2605.17493) is what exege.nn is heading for and does not implement yet.