File format · .f32
Raw 32-bit IEEE Floating Point Files: Format Facts and Recovery
How raw binary32 (FP32) floating point data is stored, why raw float files become unreadable, and how our engineers recover scientific and sensor data from them.
Raw 32-bit IEEE floating point values, known formally as binary32 in the IEEE 754 standard, are one of the most common ways computers store real numbers. You will meet them as bare data inside scientific instrument files, sensor logs, audio samples, GIS rasters, and legacy database and engineering exports. Because the values are often stored "raw", with no container or header, these files are unusually easy to corrupt and unusually easy to misread. This page explains the format, why raw float data goes wrong, and what recovery can achieve.
What raw 32-bit floating point data is
IEEE 754 defines binary32 as a 32-bit (4-byte) number made of three fields: a 1-bit sign, an 8-bit exponent, and a 23-bit stored significand, giving 24 bits of precision overall. That yields roughly 6 to 9 significant decimal digits. The largest finite value is about 3.4028235 × 10^38, and the format can represent numbers as small as roughly 1.4 × 10^-45 in denormal form.
The standard dates from 1985, and before its adoption each vendor had its own 32-bit float layout. Microsoft BASIC, for example, used a format called MBF, and IBM mainframes used a base-16 exponent. Those legacy formats are not byte-compatible with IEEE binary32, which matters a great deal in recovery: a file of old MBF floats read as IEEE values produces plausible-looking but entirely wrong numbers.
Raw float files have no magic bytes, no signature, and no self-describing header. A file of N binary32 values is simply 4×N bytes. Some writers prepend a small text or binary header; many scientific instruments write nothing at all. Endianness is not fixed by the standard, so the same values may appear little-endian on a PC acquisition system and big-endian on an instrument controller.
Where you encounter these files
Common sources include data loggers, oscilloscopes, medical imaging devices, seismic and sonar systems, weather stations, and finite-element or simulation output. Formats such as NetCDF, HDF5, and many FITS variants use binary32 as an underlying storage type, and raw float arrays appear inside GIS rasters, audio research files, and legacy Fortran or C program output. MATLAB, C, C++, Fortran, and most scientific languages read and write these values directly.
Why raw float files become unreadable
Because there is no header, nothing marks where the data begins or how it is laid out. If a header exists and is damaged, the reader loses the record length, channel count, or endianness, and the file appears as noise. Common failure patterns we see include:
- Truncated writes when an instrument loses power mid-acquisition, leaving a partial final record.
- Byte-offset shifts caused by a lost or inserted header, so every value is read from the wrong boundary.
- Wrong endianness assumptions, producing tiny denormals or absurd magnitudes.
- Legacy encodings such as MBF or VAX float misread as IEEE binary32.
- Media damage or partial overwrites that punch holes in an otherwise regular array.
- Interleaved metadata channels that a naive reader interprets as float values.
A subtle point: corruption in raw float data rarely produces an error message. Instead it produces values that look wrong, such as sudden spikes, zeros, or NaN entries, because almost any 4 bytes decode to some legal float.
What data can be recovered
Recovery is usually good. The values themselves are the payload, and each 4-byte group is independent, so damage is local. We can typically recover all records outside the damaged region, reconstruct the record structure from the data's own statistics, detect the correct endianness and offset, and flag or interpolate values that decode to NaN, infinity, or implausible outliers. Where a legacy format such as MBF is involved, we can transcode the values correctly to IEEE binary32 or to a readable text form.
How recovery works
Our engineers start by scanning the byte stream for alignment: the offset and byte order that yield a plausible distribution of floats rather than denormal noise. We then profile the values, looking for the record pitch, channel interleaving, and header boundaries. Damaged regions are mapped precisely so you know exactly which samples are lost. Finally we export the recovered array into a format you choose, such as CSV, a binary32 file with a correct header, or a structured scientific format. Where the source format is documented by the vendor, that documentation is authoritative for its layout, and we work from it whenever it is available.
Frequently asked questions
How can I tell if my file contains raw IEEE floats?
Look for a file size that is a multiple of four, and inspect a hex dump. IEEE binary32 values in real data tend to show a limited, repeating set of high bytes, whereas random bytes decode to wild exponents. A quick test is to read the file as floats in both byte orders and see which gives sensible magnitudes.
My values are all tiny or absurdly large. Is the file corrupt?
Often it is simply being read with the wrong endianness or the wrong byte offset. Both are interpretation problems rather than damage, and both are fully correctable without data loss.
Can values from an old pre-IEEE format be converted?
Yes. Formats such as MBF and VAX F-format have known layouts, and we can transcode them exactly, provided the source format can be identified from the software or hardware that wrote the file.
Is partially overwritten data recoverable?
Values outside the overwritten span are intact and recoverable. Inside the span, the original bits are gone; we can mark those samples as missing rather than guess silently.
If you have a file of raw floating point values that will not read correctly, send it to us for a free evaluation. We will assess the structure, identify the encoding, and tell you exactly what is recoverable before any repair work begins.
Send us your .f32 file for a free evaluation
Our recovery engineers will tell you exactly what can be recovered — at no cost. There is no charge if we cannot recover your data.
Start a free evaluation