FileFragment500: a documented 500-file collection for file-fragment analysis

FileFragment500 is a documented collection of 500 source files for file-fragment analysis: 50 files in each of tenrecorded-extension groups (PNG, JPG, HEIC, PDF, MP4, DOCX, PPTX, JPEG, WEBM and CSV). It comes with the 512-bytefragments, fragment pairs and similarity components derived from the files, an external test set from Govdocs1, andthe code and records used to build everything. Every file is traced to a public source and identified by its SHA-256value. The collection and its file-level split into 350 training, 75 validation and 75 test files were frozen on2026-10-27 10:47:55 UTC, before any model was trained. The 250 NapierOne files (PDF, DOCX, PPTX, CSV, MP4) are not redistributed; tools/retrieve_napierone.py downloads them from the NapierOne archives and checks every SHA-256 value. Code in scripts/ and tools/ is under the MIT licence; each source file keeps the licence of its source (ATTRIBUTION.csv).