mimetype tests files

A collection of files gathered from different sources to be used for tests that compare mimetype with the UNIX libmagic utility.

Results

TLDR: ~90% of samples identified correctly

The misidentified files, most are indeed misidentified files, but some happen because mimetype identifies more precisely than libmagic:

• XML based file formats, like GML, GPX, are seens as generic text/xml by libmagic
• mimetype identifies subtitles as text/vtt, while libmagic sees them just as plain/text
• mimetype identifies text/tab-separated-values, while libmagic sees just plain/text
• etc.

Update 01/2026: after adding the samples pronom-research the correctly identified percentage dropped from 97% to 90%.

Magika: tried magika instead of libmagic as benchmark, but magika and libmagic don't agree too much on results. It seems to be mostly magika being wrong (that is not to say libmagic is always right, all these magic number detection solutions are easy to trick). A comprehensive analysis is hard to get, but it seems libmagic is better than magika. Some key places where magika fails are:

• binary files that contain some ASCII strings: these will often be clasified as text/plain
• zip file using STORE compression (aka, no compression) containing a .docx file: will be detected as .docx

Results show the latest percentage of misidentified files and a breakdown of what are the most misidentified formats. If you want to run the tests, use these commands.

Contents

• testfiles contains all the test files (around 50 000 entries)
• zipshuffler.go reads zip files and then creates random permutations of the files inside the zip.
• truncate.go creates 3KB truncated copies of all the files
• main.go iterates over all files and compares our results with the results of file --mime