rycerzes/atarashi-dep5

Debian DEP-5 license corpus

Source-file headers labelled with the license Debian maintainers recorded for them, in the machine-readable DEP-5 debian/copyright format. No matcher was involved in producing these labels, which is what makes the corpus usable for evaluating license detectors without circularity.

Built for FOSSology/Atarashi; it is the training and evaluation set for that project's candidate-ranking model.

Not to be confused with rycerzes/atarashi-dataset

That is an earlier, unrelated set of 28,491 text fragments built for a classifier approach that was abandoned. Roughly 86% of its fragments are verbatim substrings of the reference texts they would be scored against, so recall measured on it reflects copy-detection rather than identification. Nothing here derives from it.

Contents

rows	2,050 files from 544 source packages
notice regime	671 — license prose present, answerable
no-signal regime	1,379 — no prose; the correct answer is UNKNOWN
distinct licenses (notice)	33

Fields: regime, license (SPDX-style shortname), dep5_name (Debian's own spelling, pre-normalization), text, text_sha256, package, version, path.

Labels use SPDX 3.0 naming, so GPL-2.0-only and GPL-2.0-or-later are distinct classes. Debian's notation carries the same distinction — GPL-2 means only, GPL-2+ means or-later — and collapsing them to a bare GPL-2.0, as an earlier revision did, discards something legally decisive and grades a detector leniently in both directions.

Why text is null on no-signal rows

Those files carry no license notice, so their content is plain third-party source code. Redistributing ~1,400 code excerpts from 544 packages under mixed licenses is not this dataset's to do. They ship as pointers instead — package, version, path, and the SHA-256 of the derived text — and rehydrate.py rebuilds them from Debian:

python rehydrate.py corpus.jsonl --out corpus.full.jsonl

Each rebuilt row is checked against its recorded hash, so reconstruction is verifiable and a row whose upstream has changed is reported rather than silently substituted. Notice rows carry their text inline: a license notice exists to be reproduced.

Derivation, if you would rather redo it: take the first 40 lines of the file, remove every line containing SPDX-License-Identifier:, collapse whitespace.

Label noise is ~7% — quote it with any result

Of 359 queries that two independent detectors (Atarashi and ScanCode) both answered, 48 contradict the DEP-5 label, and 26 of those have the two detectors agreeing with each other against it. Two causes:

• DEP-5 is loose about only-vs-or-later: maintainers write GPL-2 for a file whose header says "version 2 or any later version".
• Files: globs attribute a package-level license to files whose own header differs.

Good enough to compare engines against identical labels. Not an absolute accuracy ceiling, and a disagreement is not evidence of an engine error without reading the file.

How packages were chosen

Rarest-license-first. Surveying a package costs two HTTP requests; building rows from it costs about thirteen. So a broad survey picks the targets and the expensive pass is spent only where it buys a license not already covered. Random sampling saturates on MIT, GPL and Apache around 18 licenses; this reaches 34 in the notice regime and 55 overall.

The 6,000-package survey behind the selection saw 97 distinct licenses. Those 55 cover ≥92% of all license occurrences in it — license usage is extremely long-tailed, so broad coverage of the catalogue and broad coverage of reality are very different targets.

Licensing

The selection, labelling, normalization and tooling are Apache-2.0. The text of each notice row is a license notice reproduced from Debian source, under the license named in that row's license field, with full provenance in package/version/path. No other file content is redistributed.