rycerzes/atarashi-sh-blobs

Software Heritage licence blobs — prevalence and tail corpora

The labels here are ScanCode's, not ground truth. This is a distillation of one scanner's output at score ≥ 95, useful as training data and as a self-comparison regression benchmark. It cannot be used to show that anything beats ScanCode, and a result quoted from it without that caveat is misleading. The one exception is the tail config, where a human annotator reviewed and did not dispute the label — which also makes that split an easy subset for ScanCode.

Derived from the Software Heritage License Dataset (2022 edition). Built for Atarashi, FOSSology's licence identifier. Companion to rycerzes/atarashi-dep5, which carries independent labels from Debian maintainers.

Why it exists

Licence granularity. Debian maintainers write bare MIT on every MIT variant, and only 18 licences carry SPDX tags in the-stack, so no other corpus could teach a ranker MIT-advertising against MIT, or CC-BY-SA-3.0 against CC-BY-3.0. ScanCode labels every blob at full granularity, and this corpus took a model's trained licence set from 37 to 152.

Prevalence. The Software Heritage 20k sample is a random draw from 6.9M real licence blobs, so its label distribution is the real one. Most licence benchmarks take one file per licence and are macro by construction — the right instrument for "which licences can we identify at all", the wrong one for "how often are we right on real files". The two differ by a factor of two on the same engine.

Configs

config	rows	text	what it is
prevalence	15,317 label rows over 12,881 blobs	pointer	prevalence-weighted benchmark
training	1,997	pointer	licence-granularity training corpus
tail	81	inline	one human-confirmed blob per licence
baseline	12,881	—	reference predictions from a scanner run

python rehydrate.py --split prevalence attaches text to the pointer configs from the SH annex (27 MB download) and verifies every blob against its sha1. tail ships text because it was fetched one blob at a time through an API allowing 120 anonymous requests an hour.

Two traps, both of which cost real time

A blob can carry several label rows. The source CSV records one row per detected licence, so 15,317 rows cover 12,881 blobs. Group by sha1 and compare sets. Scoring rows independently grades a multi-licence file as several single-licence failures — doing that reported R@1 0.7962 for an engine whose true score was 0.9752.

Do not filter out the large blobs. Rare licences hide in multi-megabyte NOTICE and THIRD-PARTY-NOTICES aggregates; they are 22.5% of training and for licences like Unicode-TOU, FTL and NAIST-2003 they are every example. Dropping them looks obviously right and is a measured regression — capping at 50 KB moved R@1 0.9752 → 0.9600 at identical precision, because it drops 452 licences to 297.

Reference results

Measured on Atarashi's cascade, prevalence config, n=11,680 single-licence nameable blobs across 52 licences:

R@1	0.9752 ±0.0028
precision	0.9844
exact-set	0.9742
macro over 52 licences	0.7387

Scoring this pool whole found a defect four other benchmarks missed: 286 of 639 verbatim Apache-2.0 files reported as ImageMagick, whose licence is Apache-2.0 plus extra text and so wins a longest-run comparison while covering less of its own reference.

Licence and citation

CC-BY-4.0, inherited from the source dataset. The blob texts are licence files, each under its own terms. Please cite:

@article{shlicensedataset2023,
  title   = {The Software Heritage License Dataset (2022 Edition)},
  author  = {Gonzalez-Barahona, Jesus M. and Montes-Leon, Sergio and
             Robles, Gregorio and Zacchiroli, Stefano},
  journal = {Empirical Software Engineering},
  year    = {2023}, publisher = {Springer}
}
@inproceedings{zacchiroli2022licensevariants,
  title     = {A Large-scale Dataset of (Open Source) License Text Variants},
  author    = {Zacchiroli, Stefano},
  booktitle = {MSR}, year = {2022}
}

The selection, grouping, manifests and measured results are part of the Atarashi dataset pipeline.