Datasets:
license: cc-by-nc-4.0
pretty_name: 'HAIM: Human-AI Music Dataset'
size_categories:
- 100K<n<1M
task_categories:
- audio-classification
language:
- ko
- en
tags:
- arxiv:2606.01686
- music
- ai-generated-music
- ai-generated-music-detection
- plagiarism-detection
- ace-step
- suno/udio
- mureka
- MTG-Jamendo
configs:
- config_name: A1_real_MTG_audio_preview
data_files:
- split: mtg
path: A1_real_MTG_audio/mtg-*.parquet
- config_name: A2_fake_audio_preview
data_files:
- split: acestep
path: A2_fake_audio/acestep-*.parquet
- split: musicgen
path: A2_fake_audio/musicgen-*.parquet
- split: lyria_pro3
path: A2_fake_audio/lyria_pro3-*.parquet
- config_name: B_hybrid_audio_preview
data_files:
- split: B1_ai_mastered_human
path: B_hybrid_audio/B1_ai_mastered_human-*.parquet
- split: B2_human_mastered_ai_dsp
path: B_hybrid_audio/B2_human_mastered_ai_dsp-*.parquet
- split: B3_mastered
path: B_hybrid_audio/B3_mastered-*.parquet
- split: B4_professional_human_mix
path: B_hybrid_audio/B4_professional_human_mix-*.parquet
- split: B6_human_lyrics_ai_gen
path: B_hybrid_audio/B6_human_lyrics_ai_gen-*.parquet
- split: B7_variation
path: B_hybrid_audio/B7_variation-*.parquet
- split: B8_edit
path: B_hybrid_audio/B8_edit-*.parquet
- split: B9_repaint
path: B_hybrid_audio/B9_repaint-*.parquet
- config_name: C_mixing_audio_preview
data_files:
- split: C1_concat
path: C_mixing_audio/C1_concat-*.parquet
- split: C2_crossfade
path: C_mixing_audio/C2_crossfade-*.parquet
- config_name: A2_fake_links_preview
data_files:
- split: suno
path: A2_fake_links/suno-*
- split: udio
path: A2_fake_links/udio-*
- split: mureka
path: A2_fake_links/mureka-*
- config_name: B5_youtube_links_preview
data_files:
- split: b5
path: B5_youtube_links/b5-*
- config_name: A1_sonics_links_preview
data_files:
- split: sonics
path: A1_sonics_links/sonics-*
HAIM: Human-AI Music Dataset
Human-AI Music Datasets for AI Music Production Tracking Benchmark
Beyond binary "AI-or-human" — granular, role-level tracking of AI intervention across the music production workflow.
As generative platforms like Suno and Udio reach human-grade audio quality, AI now touches every stage of music production — not just full-song generation. HAIM moves detection beyond binary "AI-or-human" toward role-level tracking (Composer / Lyricist / Vocalist / Audio Engineer), with 153,686 tracks (67,000 with audio ≈ 229 GB + 86,686 link-based) across 13 categories that isolate each stage of AI intervention.
The viewer configs above are 100-row samples for preview/listening only —
load_dataset(...)returns these samples, not the full dataset. The complete data lives in the raw folders (A_full_generation/,B_hybrid/,C_mixing/) — see Download the full dataset.
Preview
In the Data Studio above, pick a config (e.g. B_hybrid_audio_preview) → a split (e.g. B1_ai_mastered_human) → press ▶ to listen. Audio configs show per-track metadata (prompt, lyrics, mode, …); *_links configs show platform/YouTube URL tables.
What is inside
A — Full generation. Fully human music (A1: MTG-Jamendo CC audio, SONICS YouTube links) versus fully AI music (A2: locally generated ACE-Step / MusicGen audio, Lyria Pro 3 via API, plus Suno / Udio / Mureka as URL manifests). The classic real-vs-fake axis, at scale and multi-platform.
B — Hybrid production. The interesting middle: AI mastering on human tracks (B1), human mastering/mixing on AI tracks (B2–B4), AI vocal covers on human tracks (B5), AI generation conditioned on human-written lyrics (B6), and AI variation / edit / repaint of real songs (B7–B9). B3/B4 explicitly model how professional human post-production can smooth out AI artifacts and evade detectors.
C — Temporal mix-sets. Human and AI segments interleaved (concat) or blended (crossfade) inside one file — for evaluating where the AI parts are on a timeline, without boundary-specific training.
Composition
| Category | Subset | Source | Tracks | Format |
|---|---|---|---|---|
| A1 · Full Human | MTG-Jamendo | CC-licensed music | 6,000 | audio |
| A1 · Full Human | SONICS | YouTube | 48,090 | links |
| A2 · Full AI | ACE-Step 1.5 | local generation | 6,000 | audio |
| A2 · Full AI | MusicGen | local generation | 5,000 | audio |
| A2 · Full AI | Lyria Pro 3 | DeepMind API | 6,000 | audio |
| A2 · Full AI | Suno / Udio / Mureka | commercial platforms | 36,556 | links |
| B · Hybrid | B1–B4 mastering & mixing | SonicMaster, FXencoder, DAW | 24,000 | audio |
| B · Hybrid | B5 AI vocal cover | YouTube | 2,040 | links |
| B · Hybrid | B6–B9 lyrics-gen, variation, edit, repaint | ACE-Step 1.5 | 8,000 | audio |
| C · Temporal | C1 concat · C2 crossfade | human+AI segments | 12,000 | audio |
Totals: 67,000 audio tracks (~240 GB) + 86,686 link-based = 153,686 tracks.
Quick start (download the full dataset)
Download the raw folders directly — not via load_dataset:
pip install -U huggingface_hub
hf download mippia/HAIM --repo-type dataset --local-dir HAIM # everything (~250 GB)
from huggingface_hub import snapshot_download
snapshot_download("mippia/HAIM", repo_type="dataset", local_dir="HAIM") # everything
snapshot_download("mippia/HAIM", repo_type="dataset", local_dir="HAIM",
allow_patterns=["B_hybrid/**"]) # one category only
- Per-folder metadata: each audio folder's
metadata.csv(file_name,track_id, + generation params where available) - Link manifests:
A_full_generation/A2_fake/{suno,udio,mureka}.json&*_tracks.csv,B_hybrid/B5_ai_vocal_human_track/b5_tracks.csv,A_full_generation/A1_real/sonics/real_songs.csv
Preview samples via datasets (not the full dataset)
from datasets import load_dataset # pip install "datasets[audio]"
# 100-row preview samples, NOT the full dataset
ds = load_dataset("mippia/HAIM", "B_hybrid_audio_preview", split="B1_ai_mastered_human")
suno = load_dataset("mippia/HAIM", "A2_fake_links_preview", split="suno")
Folder structure
├── A_full_generation
│ ├── A1_real
│ │ ├── MTG-Jamendo music subset/ # 6,000 mp3 + metadata.csv
│ │ ├── sonics/real_songs.csv # 48,090 YouTube links
│ │ └── mtg_tracks.csv
│ └── A2_fake
│ ├── acestep/ lyria-pro3/ # audio + sidecar JSON + metadata.csv
│ ├── musicgen/ # audio + metadata.csv (no sidecar)
│ └── suno[.json|_tracks.csv] udio… mureka… # link manifests
├── B_hybrid
│ ├── B1…B4, B6…B9/ # audio + metadata.csv
│ └── B5_ai_vocal_human_track/ # b5_tracks.csv + metadata.jsonl
├── C_mixing/C1_mixset_concat/ C2_mixset_crossfade/
├── A1_real_MTG_audio/ A2_fake_audio/ B_hybrid_audio/ C_mixing_audio/ # 100-row viewer samples (parquet)
├── A2_fake_links/ B5_youtube_links/ A1_sonics_links/ # 100-row viewer samples (parquet)
├── scripts/ figures/
├── DATASET_DESCRIPTION.txt
└── SUPPLEMENT_LEGAL_AND_ETHICAL.pdf
Large audio folders are sharded into d0/, d1/ subfolders (Hub 10k-files-per-directory limit).
Metadata reference (per-file columns)
A1_real/sonics/real_songs.csv — filename, title, artist, year, lyrics, duration, youtube_id, label
A2_fake/{suno,udio,mureka}_tracks.csv — platform, version, track_id, filename, page_url, audio_url
B_hybrid/B5_ai_vocal_human_track/b5_tracks.csv — track_id, original_title, url, channel, views, published, category
ACE-Step sidecar JSON (A2 acestep, B7–B9) — mode (variation/edit/repaint), concept.prompt, concept.lyrics, ace_step.result (seed, BPM, key, model version)
B6_metadata.csv — human K-pop lyrics used for conditioning (text only, no original audio; ref lt_dataset): LID, ARTIST, EN_TITLE, KR_TITLE, IS_OFFICIAL, URL
Pipeline
Citation
@article{go2026haim,
title={HAIM: Human-AI Music Datasets for AI Music Production Tracking Benchmark},
author={Go, Seonghyeon and Kim, Yumin},
journal={arXiv preprint arXiv:2606.01686},
year={2026}
}
License
Data: CC-BY-NC-4.0, non-commercial academic research only — commercial-platform content is released as URL manifests only per platform ToS. Code: MIT (GitHub). Please credit MTG-Jamendo and SONICS per their licenses. Full analysis: SUPPLEMENT_LEGAL_AND_ETHICAL.pdf