Data

manyLatents provides synthetic manifold datasets for benchmarking and a precomputed loader for custom data.

dataset	config	key params
Archetypal	data=archetypal	n_components=3, simplex_radius=1.0, n_obs=5000
DLATree	data=dla_tree	n_dim=50, n_branch=10, branch_lengths=[300, 300, 25, 300, 300, 300, 25, 300, 300, 300]
DLATreeFromGraph	data=dla_tree_from_graph	n_dim=100, sigma=0.5, mode=full
DLATreeFromGraph	data=dla_tree_from_graph_nogaps	n_dim=100, sigma=0.5, mode=full
GaussianBlob	data=gaussian_blobs	n_samples=1000, n_features=2, centers=5
MHI	data=mhi_split	cache_dir=${cache_dir}, mmap_mode=None, mode=split
HFText	data=pile	dataset_name=None, dataset_config=None, data_dir=/network/datasets/pile
Precomputed	data=precomputed	path=None, label_col=None, mode=full
ReasoningTrace	data=reasoning_trace	trace_store_path=None, tensor_key=pooled_steps, layer_index=-1
SaddleSurface	data=saddle_surface	n_distributions=100, n_points_per_distribution=50, noise=0.1
SwissRoll	data=swissroll	n_distributions=10, n_points_per_distribution=100, noise=0.2
Torus	data=torus	n_points=1000, noise=0.1, major_radius=5.0
HFText	data=wikitext	dataset_name=wikitext, dataset_config=wikitext-2-raw-v1, tokenizer_name=gpt2

Domain-specific datasets (genomics, single-cell) are available via the manylatents-omics extension.

Precomputed Data

Load your own data from .npy or .npz files:

uv run python -m manylatents.main data=precomputed data.path=/path/to/data.npy algorithms/latent=umap

Sampling

Large datasets are subsampled before metric evaluation. Configure under metrics.sampling:

strategy	config	defaults
FarthestPointSampling	sampling/farthest_point	seed=42, fraction=0.1
RandomSampling	sampling/random	seed=42, fraction=0.1
StratifiedSampling	sampling/stratified	stratify_by=population_label, seed=42, fraction=0.1