Find files, editable templates and browser test targets by what you need to make or test. The directory below is cut by format; the two collections under it cut the same library by subject and by workflow.
Page 5 of 6; 24 results per page.
The real FFT of a 64-sample sine stored as complex128, where each element is an interleaved pair of doubles rather than two separate planes. Readers without a complex type usually flatten it, which doubles the reported length and shifts the peak bin.
Five timestamps stored as datetime64 with second resolution, where the unit is part of the dtype descriptor and the values on disk are plain int64 epoch seconds. Dropping the unit turns 2026-01-01 into the integer 1767225600 without any complaint.
Every structurally interesting float16 value in one array: one plus epsilon, the largest finite value, the smallest normal and subnormal, negative zero, infinity and a NaN. It is the compact half-precision counterpart to the float32 and float64 boundary fixtures.
The identical temperature field carried by the CF NetCDF and HDF5 fixtures, stored as a bare float32 .npy. Comparing the three shows exactly what a plain array container loses: the numbers survive, the units and axes do not.
The baseline .npy fixture: a 12x8 float64 array in version 1.0 format with a little-endian descriptor and the header padded to the mandatory 64-byte alignment. Every other array in this family varies exactly one property away from it.
Seven doubles written big-endian, including negative zero and values at both ends of the exponent range. Byte-swapping a double is unrecoverable by inspection. Swap pi and you get 3.2e-192, which looks like a plausible tiny number rather than an error.
The same 4x6 matrix stored column-major with fortran_order set to True in the header. Ignoring that one boolean does not produce an error, it produces a transposed matrix: the quietest failure in the whole .npy format.
The same nine int16 values written big-endian, so the header descriptor reads '>i2'. A loader that ignores the descriptor and assumes native little-endian order returns 13330 where the file says 4660, without any error.
Nine int16 values including both type extremes, stored little-endian so the descriptor in the header reads '<i2'. It is one half of an endian pair whose values are identical and whose bytes are not.
A 16-record structured array carrying a value, its uncertainty and a quality flag, with the physical unit encoded in the field-name suffix because .npy has nowhere else to put it. The header descriptor is a list of field tuples rather than a single type string, which is the parsing branch this file exercises.
A rank-0 .npy holding a single double, whose header shape is the empty tuple rather than (1,). Indexing it with [0] raises rather than returning the value, so it separates parsers that model rank properly from parsers that assume at least one axis.
A valid .npy file with a full header, a declared (0, 4) shape and a data section of zero bytes. It preserves the column count across an empty result, and it separates readers that model emptiness from readers that treat it as failure.
The identical five arrays bundled with deflate instead of stored, which is the only difference between NumPy's savez and savez_compressed. Loading both and comparing proves a reader ran the decompressor rather than trusting stored-size shortcuts.
An npz-style bundle: a ZIP whose members are .npy files, which is exactly what NumPy's savez produces. np.load opens it by ZIP magic rather than by extension, so the arrays load straight from this .zip; rename it to .npz for tooling that checks the suffix.
The stored integer, the unpacked value, the original float and the quantisation error for all 120 cells of the packed NetCDF grid. It makes the cost of int16 packing explicit and gives an unpacking implementation an exact per-cell target.
Three synthetic protein sequences that begin with methionine and include the X, B and Z ambiguity codes, which are legal IUPAC residues rather than errors. A validator restricted to the twenty standard amino acids rejects all three files.
The same sixteen doubles written in network byte order, so every eight-byte group is the mirror of the little-endian twin. Decoding it with the wrong assumption produces small plausible-looking numbers rather than an error, which is why this pair exists.
Sixteen doubles as a bare 128-byte block with no header, so byte order is knowledge the reader has to bring rather than something the file declares. It is one half of an endian pair that only a correct assumption tells apart.
Every value in the raw endian pair with its exact byte sequence in both orders, so a decoder can be checked offset by offset instead of value by value. Negative zero and the subnormal entries are the rows where a byte-order mistake stops looking like a mistake.
Six alignments against a synthetic contig with a full @HD/@SQ/@RG/@PG header, both strands via the reverse flag, and CIGAR strings covering match, soft clip, insertion and deletion. Reference span and query length come apart in three of the records, which is where CIGAR arithmetic gets tested.
Each alignment's CIGAR decomposed into how many query bases and how many reference bases it consumes, with the resulting 1-based reference end. The three records where those numbers differ are the whole reason CIGAR arithmetic is worth testing.
Twelve doubles printed at 15, 16 and 17 significant digits alongside their shortest round-tripping form, with a column recording which precisions actually recover the original bits. Fifteen digits is the default in a lot of formatting code and it loses values, which this table demonstrates rather than asserts.
Eight measurements whose unit symbols use the codepoints that have Unicode look-alikes: MICRO SIGN against GREEK MU, OHM SIGN against GREEK OMEGA, ANGSTROM SIGN against A-with-ring. Each row also carries an ASCII fallback, so a normalisation bug shows up as a column mismatch rather than as an invisible change.
Nine ways of writing zero, what each parses to, and what dividing one by the result gives. Negative zero compares equal to positive zero yet behaves differently in division and in sign propagation, so a formatter that drops the sign changes results without changing any comparison.
