v2/README.md

metadata

license: cc-by-nc-4.0
task_categories:
  - audio-classification
language:
  - en
pretty_name: ArtifactBench v2
size_categories:
  - n<1K

ArtifactBench v2

ArtifactBench v2 is a metadata-first, lineage-aware evaluation suite for AI-generated music detectors. The primary frozen protocol contains 828 tracks: 605 AI-generated and 223 real. It has 166 calibration, 83 validation, and 579 sealed-test rows.

What is released

• path- and title-free track IDs derived from content digests;
• source and generator strata;
• SHA-256 content identities and lineage IDs;
• frozen calibration, validation, and test assignments;
• retrieval references where redistribution is appropriate;
• version-pinned runner code, raw model probabilities, structured failures, aggregate tables, and confidence intervals.

No new audio bundle is released with v2. Source data must be acquired from its upstream provider under that provider's license and terms. Suno-, Udio-, and YouTube-referenced audio is not redistributed.

Intended use

The suite is intended for detector comparison, reproducibility checks, source- shift analysis, and research on evaluation protocol. It is not evidence of authorship, copyright status, fraud, or the degree of AI assistance in a work.

Protocol

Thresholds are selected on calibration data only by maximizing TPR subject to FPR no greater than 5%. Validation is diagnostic and cannot change a threshold. The sealed test is evaluated once. Results include AUROC, AUPRC, F1, balanced accuracy, TPR/FPR, per-source rates, lineage-bootstrap intervals, and inference coverage. Pairwise comparisons use the common successfully scored track-ID set. The split unit is lineage_id, stratified by source with the published salt artifactbench-v2-primary-20260920; split_rank_sha256 allows the frozen assignment to be verified without exposing the private lineage key.

ArtifactNet evaluates seven evenly spaced four-second chunks. The declared v2 policy reports the median of finite scores when at least four chunks are valid; otherwise the track is an inference failure. A strict any-non-finite sensitivity result and the chunk-level failure log are included in the result bundle.

Frozen result snapshot

On the 562-track common-success test intersection, the calibration-selected results are:

Model	AUROC	AUPRC	F1	Balanced accuracy	TPR	FPR
ArtifactNet v9.4	0.982	0.994	0.923	0.918	0.865	0.029
Deezer ISMIR	0.761	0.925	0.758	0.776	0.624	0.072
SpecTTTra alpha	0.299	0.688	0.159	0.526	0.087	0.036
CLAM	0.284	0.647	0.000	0.500	0.000	0.000

CLAM's calibration-selected threshold is the next representable value above 1.0 because lower thresholds cannot satisfy the 5% calibration-FPR constraint; its native 0.5 threshold has TPR 0.889 and FPR 0.942 on the same intersection. This degenerate operating point must therefore be read together with AUROC and native-threshold results. ArtifactNet scores 562 of 579 sealed-test attempts; the other three models score all 579. Treating every ArtifactNet failure as an adversarial class error gives balanced accuracy 0.865. Exact values, bootstrap intervals, raw probabilities, and structured failures are in results/frozen_protocol_metrics.json.

Evaluated baselines

• ArtifactNet v9.4 ONNX revision e915f0dc5962a57536bbe1f78b66adcb48dfae4c;
• SpecTTTra alpha 120-second revision 094b32a5545098a71c113f6ae9d5c55310564268;
• Deezer ISMIR 2025 research detector revision d2180598fed79e3f917e8050a00439982466e5c6;
• CLAM repository revision 74e3a3277e1dfe9ae9ed433b6e8c51d74e9e1d9b, with checkpoint and encoder digests recorded in the run manifest.

The frozen inference runner revision used for this snapshot is 55444b33b1725e6f69905716b02e6a881d7c8a07. Reproducible analysis, independent metric verification, public-result sanitation, and release-audit tooling were added in release-candidate revision 1a4ce7ffae7513dc6ca97380d57c983ee56028ac.

Limitations

The cohort is modest and concentrated in selected generator families. Some platform version labels are incomplete. Public training lineage cannot resolve all CLAM real-track identities. Cryptographic hashes do not detect every perceptual derivative. The binary task does not cover hybrid human--AI production.

Licensing

Benchmark-authored metadata and result files are released under CC BY-NC 4.0. Runner code is released separately under the repository's MIT License. These licenses do not grant rights to upstream audio, model weights, platform content, or third-party metadata. See LICENSES.md and fma_track_licenses.json before use.