CLI reference¶

class-e is one flat command line with nine commands. Run class-e COMMAND --help for the full flag list.

Command	Purpose
`generate`	compute training and test sets with CLASS
`merge`	merge the chunks of a design
`status`	progress of chunked generation
`train`	fit emulators and save a bundle
`evaluate`	accuracy report of a bundle on a test set
`info`	describe a bundle
`sample`	bundle preflight and single-point prediction
`stitch`	join bundles covering adjacent $\ell$ ranges
`inspect`	dataset notebook (needs the `ui` extra)

Most commands read a YAML configuration; --set KEY=VALUE overrides any value with a dotted path.

generate¶

class-e generate configs/lcdm.yaml -N 100 -o TT -o TE -o EE
class-e generate configs/w0wacdm.yaml --workers 2 --threads 4
class-e generate configs/ede-axiclass.yaml --classy ~/github/AxiCLASS
class-e generate configs/w0wacdm.yaml --slurm --submit

Flag	Meaning
`--design NAME`	one design of `samples.designs` (default: all)
`--out PATH`	output file of one design (with `--chunk`: directory of the parts)
`-N`, `--N`	number of cosmologies
`-m`, `--method`	`sobol` (use $N = 2^m$), `lhc`, `lhs`, `cov`
`--cov PATH`	covariance matrix for `cov` (implies it)
`--seed`	design seed (overrides `samples.seed`)
`-o`, `--observable`	observable to compute, repeatable
`--lmin`, `--lmax`, `--kmin`, `--kmax`	multipole range and $P(k)$ grid [h/Mpc]
`--workers`	chunks computed at once (an integer or `auto`)
`--threads`	OpenMP threads per CLASS run
`--n-chunks`, `--chunk`	split a design; compute one chunk
`--checkpoint-every`	checkpoint frequency in points
`--classy PATH`	CLASS build; default `$CLASS_E_CLASSY`, then `theory.classy.path`, then the installed classy (details)
`--no-isolate`	compute in this process (debugging)
`--slurm`, `--submit`	write, and optionally submit, SLURM jobs
`-d`, `--dry-run`	print the settings and exit
`-f`, `--force`	overwrite without asking
`-i`, `--interactive`, `-e`, `--edit`	Marimo notebook (edit mode)

Output is .npz (format v2) plus a .meta.json sidecar, see Datasets and bundles.

merge¶

class-e merge data/toy/parts/toy_train_N256_s0 -o data/toy/data_train.npz

PREFIX is everything before .partXXXofYYY.npz. --max-failure-fraction sets the accepted CLASS failures over the whole design.

status¶

class-e status data/toy/parts [--json]

Finished chunks, checkpointed points, failures and running chunks per design.

train¶

class-e train configs/toy-pipeline.yaml
class-e train --dataset data.npz --backend jax --N-PCA 20 --out models/lcdm.cemu

Flag	Meaning
`--dataset`, `--test-dataset`, `--test-fraction`	training set; separate test set, or the held-out fraction (0.2)
`--backend`	`sklearn` (default), `jax` or `vae`
`--N-PCA`	PCA components (latent size for `vae`; 0 disables PCA, sklearn only)
`--kernel`, `--ard / --no-ard`	GP kernel (`rbf`, `matern32`, `matern52`, `rq`), one length scale per input
`--optimizer`, `--n-restarts`	`auto`, `lbfgs` or `adam` (JAX and VAE); random restarts
`--log-space / --no-log-space`	emulate $\log Y$
`--standardize / --no-standardize`	standardize inputs, see Backends
`--input NAME=EXPR`	train on an expression of the parameters, repeatable
`--out PATH`	save the `.cemu` bundle
`--learning-curve N1,N2,...`	instead of training a bundle: test accuracy when training on the first N points
`--pca-curve K1,K2,...`	instead of training a bundle: test accuracy per number of PCA components
`--plots DIR`	with the curves: write the figure (needs `class-e[plot]`) and the numbers (JSON)
`--x64 / --no-x64`	float64 for JAX backends
`-i`, `--interactive`, `-e`, `--edit`	Marimo notebook (needs the `ui` extra)

evaluate¶

class-e evaluate bundles/toy.cemu --config configs/toy-pipeline.yaml \
    --plots plots/toy --report plots/toy.html
class-e evaluate model.cemu -t test.npz --ell-ranges 2,30,500,2500 --by "w0 + wa:-3,-0.5,0"

Flag	Meaning
`-t`, `--test`	test set `.npz`
`-c`, `--config`	pipeline YAML (its `train.test_design` and `evaluate:` section)
`-o`, `--observable`	observables to evaluate, repeatable
`--ell-ranges`	multipole range edges
`--by "EXPR:E1,E2,..."`	errors per bin of a parameter expression
`--worst N`	list the worst test points (default 5)
`--json FILE`, `--report FILE`	JSON and self-contained HTML report
`--plots DIR`	accuracy figures (needs `class-e[plot]`)

CMB errors are in units of the cosmic variance, others are relative. --plots draws one figure per kind of observable: bands (CMB spectra against ℓ, in percent and in σ_CV), relative_bands (Hz, chi against 1+z, mPk against k) and scalars (H0, rs_drag, ... : the sorted relative error over the test points).

info¶

class-e info bundles/toy.cemu

Backend, compression, standardization and differentiability of each emulator. Without a path, models/emulators.cemu is used. A published emulator's name (class-e info lcdm@jax) is downloaded first; see Published emulators.

diff¶

class-e diff a.cemu lcdm@sklearn

Compares two bundles: their parameter space (fingerprint, CLASS build, settings), inputs, outputs (observables, backends, grids), training (dataset, metrics) and software versions. See Parameter space and CLASS build.

list, download, verify, lineage, reproduce, publish, registry, cache¶

class-e list [--installed | --remote] [--registry NAME] [--offline]
class-e download data lcdm [--split test] | --for lcdm@jax [--yes]
class-e lineage lcdm@jax
class-e reproduce lcdm@jax -o again.cemu
class-e publish FILE --space lcdm --registry-file registry.json [--to file|github|hf ...] [--dry-run]
class-e release --registry-file registry.json [--author "Family, Given"] [--sandbox] [--dry-run]
class-e search [-p NAME[=LO:HI]] [-o OBS] [--class NAME] [--kind emu|data] [--json]
class-e registry init DIR | check FILE [--download emu|all] [--reachable] [--strict]
class-e verify lcdm@jax
class-e download emu lcdm [--backend jax] [--version 2] [--registry official] [--force]
class-e registry list | add NAME LOCATION | remove NAME | refresh
class-e cache path | remove NAME | prune

Published emulators: what the registries list and what is downloaded, downloading one (checked against its sha256, kept in the cache), the registries read, and the cache. See Published emulators.

sample¶

class-e sample model.cemu -o TT -p h=0.674 -p omega_b=0.0224 -p omega_cdm=0.12

A preflight, not a sampler: it validates a bundle and optionally predicts a single point (-p NAME=VALUE, repeated). Posterior sampling is not implemented in the CLI.

stitch¶

class-e stitch low.cemu high.cemu -o full.cemu --overlap 100

Joins bundles on adjacent $\ell$ ranges, blending smoothly over the overlap (-f overwrites).

inspect¶

class-e inspect [-e]

Opens the dataset-inspection notebook.

Legacy pickles

info and sample refuse pickled emulators unless --allow-pickle is given: loading a pickle executes code from the file.