NPY Array Decoder
Paste a NumPy .npy file as base64 or hex and read its dtype, shape, order and byte layout — then export the values as JSON or CSV. Versions 1.0, 2.0 and 3.0; integers, floats, bools, complex and fixed-width text; C and Fortran order. Runs entirely in your browser, no upload, no sign-up.
About this tool
.npy is NumPy's own on-disk format for a single array: a short magic number, a
version, a Python-dict header holding the descr (dtype), fortran_order and
shape, then the raw element bytes with no compression and no padding. It is
trivial for NumPy to read and completely opaque to everything else — open one in
a text editor and you get a line of readable header followed by binary noise.
This decoder reads the format directly. Paste the file's bytes as base64 or hex and it reports the dtype, the shape, whether the data is stored row-major (C) or column-major (Fortran), where the data starts, and how many bytes it occupies — then renders the values as a readable report, as JSON, or as CSV you can drop into a spreadsheet. Nothing is uploaded: the parser is compiled to WebAssembly and runs in this page.
To get the bytes out of a file, base64-encode it — base64 -w0 array.npy on
Linux, base64 -i array.npy on macOS, or
certutil -encode array.npy out.txt on Windows — then paste the result. Hex
works too (xxd -p array.npy), and Input encoding: auto tells the two apart
on its own.
Worked example
A 2x3 array of doubles saved with numpy.save is 176 bytes: a 10-byte
prologue, a 118-byte header padded out to a 64-byte boundary, then 6 x 8 = 48
bytes of data. Pasting its base64 with Output: summary gives:
NumPy .npy file, format version 1.0
dtype: float64 (descr <f8, 8 bytes per element, little-endian)
shape: (2, 3) - 2 dimensions, 6 elements
order: C (row-major)
layout: header 118 bytes, data starts at offset 128, data 48 bytes
values: all 6 elements
[[1, 2, 3.5], [4, 5, 6]]
Switch Output to csv and the same file becomes 1,2,3.5 / 4,5,6; switch
it to json and you get the metadata plus a nested data array; switch it to
header and you get the metadata alone, with no values — useful when you only
want to know what is in a large file.
What it reads
- Format versions 1.0, 2.0 and 3.0 (the versions differ in the width of the header-length field and the header's text encoding).
- dtypes —
bool,int8/16/32/64,uint8/16/32/64,float16/32/64,complex64/128, fixed-width bytes (S<n>) and fixed-width text (U<n>), little-endian or big-endian.float16is widened by hand, so half-precision arrays decode exactly, subnormals included. - Both memory orders — Fortran-ordered (column-major) data is re-indexed to row-major before it is printed, so the values always read in the same order NumPy would print them.
- Any rank — 0-d scalars, vectors, matrices and higher-rank arrays; empty arrays (a zero in the shape) are reported rather than treated as an error.
Limits and edge cases
- 8 MiB of decoded file bytes maximum — every element is materialised in memory before rendering.
- 1000 values are rendered by default; raise Max values rendered up to
100000. Past that cap the array is truncated in row-major order,
jsonandsummaryswitch to a flat list and settruncated: true, andcsvemits whole rows only. - Not supported, by design: object (
O) arrays, which hold pickled Python objects and cannot be read without executing code; structured/record dtypes, whosedescris a list of named fields;datetime64/timedelta64; and extended-precision (float96/float128) values, which are platform specific. Each is rejected with a message that says which one it hit. .npzfiles are ZIP archives of.npymembers, not.npyfiles — unzip first and decode a member.- Complex numbers and non-finite floats (
NaN,Infinity) have no JSON literal, so injsonoutput they are emitted as strings; the JSON always parses.
FAQ
How do I turn my .npy file into something I can paste here?
Base64-encode it. On Linux base64 -w0 array.npy, on macOS base64 -i array.npy,
on Windows certutil -encode array.npy out.txt (then strip the BEGIN/END lines),
or in Python import base64, pathlib; print(base64.b64encode(pathlib.Path("array.npy").read_bytes()).decode()).
Hex from xxd -p array.npy works just as well. Leave Input encoding on
auto and the tool works out which one you pasted from the file's own magic
bytes — a .npy file always begins with byte 0x93, so hex starts 93 and
base64 starts k05VTVBZ. A data:application/octet-stream;base64, prefix is
accepted and ignored.
Why does it refuse my file with "object arrays hold pickled Python objects"?
The array was saved with dtype=object — the elements are not numbers but
pickled Python objects, and the only way to turn them back into values is to run
the pickle, which can execute arbitrary code. No decoder should do that to a file
you pasted from somewhere else. Re-save the array with a concrete numeric dtype
(arr.astype("float64"), say) and it will decode. The same reasoning applies to
structured/record dtypes: their descr is a list of named fields rather than one
dtype, so save the fields as separate arrays first.
What does "truncated .npy data" mean?
The header and the data disagree. The header declares a dtype and a shape, which
together fix exactly how many bytes must follow — 2 x 3 float64 values need
48 — and fewer bytes than that are present. The error prints both numbers so you
can see the size of the shortfall. In practice it means the paste was cut off,
the file was copied while it was still being written, or only part of a larger
file was captured. Extra bytes after the array are harmless: they are reported
as unused trailing bytes rather than treated as an error.
Why is my Fortran-ordered array printed in a different order than the file?
Because the file stores it column-major and this tool prints row-major, which is
how NumPy itself displays an array regardless of how it is laid out in memory. A
2x3 array stored as 1 4 2 5 3 6 on disk is shown as [[1, 2, 3], [4, 5, 6]].
The summary and header outputs both state fortran_order, so you can always
see how the bytes were actually written.
How does CSV output map an array with more than two dimensions?
The last axis becomes the columns and every earlier axis is folded into the rows,
so a 2x3x4 array comes out as 6 rows of 4 values, in row-major order. A 1-d
array is written one value per line, matching numpy.savetxt, and a 0-d scalar
is a single value. There is no header row — the array has no column names — and
fields containing the delimiter, a quote or a newline are quoted the RFC 4180
way. Set CSV delimiter to ;, |, tab or any single character.
Developer & Automation Access
Run it from the terminal
Same engine as this page, headless — via the gizza CLI:
gizza tool npy-array-decoder "k05VTVBZAQB2AHsnZGVzY3InOiAnPGY4JywgJ2ZvcnRyYW5fb3JkZXInOiBGYWxzZSwgJ3NoYXBlJzogKDIsIDMpLCB9ICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgIAoAAAAAAADwPwAAAAAAAABAAAAAAAAADEAAAAAAAAAQQAAAAAAAABRAAAAAAAAAGEA="New to the CLI? Get gizza →
Open it by URL
Pre-fill and auto-run this tool with query parameters — the names match the API/CLI:
https://gizza.ai/tools/npy-array-decoder/?input=k05VTVBZAQB2AHsnZGVzY3InOiAnPGY4JywgJ2ZvcnRyYW5fb3JkZXInOiBGYWxzZSwgJ3NoYXBlJzogKDIsIDMpLCB9ICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgIAoAAAAAAADwPwAAAAAAAABAAAAAAAAADEAAAAAAAAAQQAAAAAAAABRAAAAAAAAAGEA%3D&input_format=auto&output=summary&limit=1000&delimiter=%2CMachine-readable descriptor: tool.json — title + parameters JSON Schema for agents.