Science and machine learning

npy npz safetensors gguf fits mat nii nrrd zarr grib mtx, and genomics: fasta fastq bed gff3 gtf vcf sam bam

Arrays, model weights, sky images, brain scans, weather grids and sequencing reads open as what they hold, read on your machine — no Python, no notebook to set up first. Which reader a file needs, its first bytes decide.

Arrays and model weights

File	Opens as
NumPy npy	its elements as rows, a page at a time at any size; a picture first when it is one
NumPy npz	a folder of its arrays, each opening as an npy
safetensors	a table of its metadata and tensors: name, type, shape, elements, bytes and offset
llama.cpp's gguf	the same table: the model's settings, then every tensor, quantized ones named by their type
MATLAB mat (version 4 and 5)	a sheet a variable: numbers, text, sparse matrices, cells and structs
Matrix Market mtx	a row per entry — row, column, value — or the dense grid

A 2-D NumPy array is a row per row and a column per column; a structured one, a column per field. An array shaped like a picture — height × width × 1, 3 or 4 bytes — opens as that picture, its rows next. An npy also opens in a notebook, for SQL over its elements.

A safetensors or gguf file is read only as far as its header: the weights are never loaded, so a model of many gigabytes opens at once. A shard of a split model says which one it is — "2 of 5". A long metadata list, like a tokenizer's vocabulary, shows its count and first values.

A MATLAB 7.3 file is HDF5 underneath and is not read as a mat; it says so.

Images and volumes

File	Opens as
FITS fits fit fts	its first image, with the header cards beside it; every part (HDU) as a sheet
NIfTI nii nii.gz	the scan, slice by slice; the voxels and the header as tables
NRRD nrrd nhdr	the same, the data read from the file the header names beside it
Zarr zarr (version 2, zipped)	the array, slice by slice
GRIB grib grb grib2	its first parameter, longitude by latitude, a slice for each of its fields

A volume opens on its middle slice. Axial, Coronal and Sagittal pick the plane, and a slider steps through each further axis — the slice, then the time. ← → step a slice, ↑ ↓ the time, PageUp PageDown a tenth of the way, Home End the first and the last. Every slice uses the same grey range, so two slices compare.

A FITS file is a series of header-and-data units (HDUs); a table among them — binary or ASCII — is a sheet with a column per field. GRIB editions 1 and 2 are read, simple packing and JPEG 2000 among their encodings; the file's other messages are listed in its facts.

Genomics

Every one of these opens as a table, a row per record, its text a click away:

File	A row for each, with
FASTA fasta fa fna faa	sequence: identifier, description, length, sequence
FASTQ fastq fq	read, with its quality and the mean quality
BED bed	region, in UCSC's column names, BED3 to BED12
GFF3 and GTF gff3 gff gtf	feature: the nine columns, and a column for each attribute
VCF vcf	variant: the #CHROM columns and one per sample
SAM sam, BAM bam	alignment: the eleven fields and a column per optional tag

A record that is broken still shows as a row, with a last column saying what is wrong. A BAM's structure mode lays out its header, references and compressed blocks. A .vcf that is a contact card opens as one — see records and trees.

HDF5 and NetCDF files open in a notebook, a row for each dataset — see HDF5 and NetCDF.

Next: databases.