Copyright analysis focuses on the similarity between two works. However, generative AI models learn from training data through functional relations beyond similarity. Training Data Attribution (TDA) estimates the influence of each training instance on a generated output and may serve as a basis for remuneration. However, prior studies find global influence, where a few training instances rank as the most influential for almost every generation. Used directly, TDA scores would distribute part of every remuneration to these instances regardless of what was generated. To address this, we propose a new approach in which TDA scores are decomposed into three levels of influence, namely global indicating the influence of a training data to a trained model, group for its influence on a group of generated outputs (that could be clustered together), and specific for its influence on a single generated output. Our study finds that, across several music generative models and TDA methods, global- and group-level influence is non-negligible in every case. Based on these observations, we propose the term functional dependency for the relation of training data to a trained model or to its outputs. This notion both encompasses and distinguishes dependencies of different scopes, allowing TDA scores to be interpreted as distinct forms of contribution in legal, economic and ethical discussions on remuneration.
Figures & tables
Figure 1. Decomposition of the standardized score matrix into the three levels.
Model
MusicGen ( Copet et al., 2023 )
InspireMusic ( Zhang et al., 2025 )
MusicGen-S
MusicGen-L
InspireMusic-Base
InspireMusic-1.5B
Fine-tuned Parameters
0.4B
3.3B
0.5B
1.5B
Frozen Components
T5 text encoder ( Raffel et al., 2020 ) , EnCodec ( Défossez et al., 2023 )
WavTokenizer ( Ji et al., 2025 ) , SRFM ( Zhang et al., 2025 )
Training Data
30,000 ten-second segments from 10,000 MTG-Jamendo tracks ( Bogdanov et al., 2019 )
Generation
T=480 , from 8 prompts for each of 12 genres with m=5 random seeds
TDA Methods
GradDot ( Charpiat et al., 2019 ) , Grad-Cos ( Hanawa et al., 2021 ) , TracIn ( Pruthi et al., 2020 ) , LoGra ( Choe et al., 2025 ) , FactGraSS ( Hu et al., 2025 ) , Trak ( Park et al., 2023 ) ( λ=1 ), and the Trak variants λ=0.1 , λ=0.01 and ensemble
Table 1. Experimental setup.
Figure 2. Distribution of genre tags in the whole MTG-Jamendo dataset (55,525 tracks) and in the fine-tuning dataset (10,000 tracks), for the 25 most frequent genre tags. A track can have several genre tags, so the percentages do not sum to 100.
global
group
Method
S
L
IM
IM-1.5B
S
L
IM
IM-1.5B
Random
0.002
0.002
0.002
0.002
0.02
0.02
0.02
0.02
GradDot
0.106
0.131
0.360
0.403
0.212
0.151
0.113
0.118
Grad-Cos
0.106
0.128
0.359
0.401
0.209
0.148
0.113
0.119
TracIn
0.135
0.139
0.454
0.651
0.239
0.214
0.128
0.173
LoGra
0.059
0.070
0.132
0.134
0.073
0.070
0.076
0.101
Table 2. Proportion of the global and group levels ( group by TDA score clusters, k=12 ). S, L, IM and IM-1.5B denote MusicGen-S, MusicGen-L, InspireMusic-Base and InspireMusic-1.5B.
Figure 3. Results of the group -level deletion test for Trak with λ=1 in MusicGen-S (left) and InspireMusic-Base (right). Each row is the removed deletion set of a TDA score cluster, and each column is the generated outputs of a TDA score cluster. Each entry is the mean ΔLL ( ×103 ) over the generated outputs of the column and three retraining seeds. Outlined cells are the diagonal.
MusicGen-S
MusicGen-L
InspireMusic-Base
InspireMusic-1.5B
∥d∥
0.208
0.192
0.290
0.282
Method
ρalign
ρnorm
ρalign
ρnorm
ρalign
ρnorm
ρalign
ρnorm
GradDot
1.00
−0.06
1.00
−0.06
1.00
+0.14
0.99
+0.15
Grad-Cos
1.00
−0.06
1.00
−0.05
1.00
+0.14
1.00
+0.14
TracIn
0.81
+0.04
0.70
+0.05
0.90
+0.32
0.85
+0.31
LoGra
0.61
−0.07
0.46
−0.09
0.71
+0.11
0.44
+0.17
Table 4. Norm of the common direction ∥d∥ , and Spearman correlations ρalign and ρnorm of ziglob .
Trak
Trak
Trak
Trak
Model
GradDot
Grad-Cos
TracIn
LoGra
FactGraSS
λ=1
λ=0.1
λ=0.01
(ensemble)
Random
Genre
MusicGen-S
0.103
0.104
0.118
0.166
0.201
0.149
0.141
0.139
0.143
0.081
MusicGen-L
0.102
0.100
0.109
0.190
0.194
0.137
0.128
0.122
0.137
0.081
InspireMusic-Base
0.102
0.116
0.107
0.163
0.179
0.175
0.161
0.161
0.168
0.081
InspireMusic-1.5B
0.107
0.107
0.090
0.122
0.121
0.115
0.111
0.104
0.115
0.081
Table 5. Wallace coefficient W between the TDA score clusters and the feature clusters of each criterion.
Figure 4. Generated outputs of MusicGen-S in the CLAP embedding space with UMAP, coloured by genre (left) and by the TDA score clusters of GradDot (right). A colour of a TDA score cluster is only a cluster label and does not indicate a genre.
Trak
Trak
Trak
Trak
Model
GradDot
Grad-Cos
TracIn
LoGra
FactGraSS
λ=1
λ=0.1
λ=0.01
(ensemble)
MusicGen-S
0.55
0.56
0.55
0.22
0.20
0.30
0.30
0.30
0.30
MusicGen-L
0.50
0.52
0.54
0.22
0.20
0.29
0.30
0.30
0.29
InspireMusic-Base
0.49
0.51
0.51
0.27
0.27
0.32
0.31
0.32
0.30
InspireMusic-1.5B
0.53
0.57
0.51
0.36
0.35
0.33
0.33
0.33
0.34
Table 6. Fraction of change in the 50 training instances with the highest specific level scores when the criterion changes from genre to the TDA score clusters.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Item
MusicGen-S
MusicGen-L
InspireMusic-Base
InspireMusic-1.5B
Released model
facebook/musicgen-small
facebook/musicgen-large
iic/InspireMusic (Base)
iic/InspireMusic-1.5B
Trained part
decoder (419.6M)
decoder (3,255.4M)
LLM (501.6M)
LLM (1,556.6M)
Frozen
T5 text encoder ( Raffel et al., 2020 ) , EnCodec ( Défossez et al., 2023 )
WavTokenizer ( Ji et al., 2025 ) , flow-matching model
zt−zglob over the 30,000 training instances, L2-normalized
k -means
Genre
first tag of the prompt
12 genres, 40 generated outputs each
MFCC ( Davis and Mermelstein, 1980 )
mean and variance over time of 13 MFCCs (26 dimensions), standardized over the generated outputs
k -means
Log-mel spectrogram ( Müller, 2015 )
mean over time of the log-mel spectrum (128 bands, dB), standardized over the generated outputs
k -means
Chroma ( Müller, 2015 )
mean and variance over time of the CQT chroma (24 dimensions), standardized over the generated outputs
k -means
Appendix
Table 9. Features and grouping of each criterion.
Item
MusicGen-S
MusicGen-L
InspireMusic-Base
InspireMusic-1.5B
GPU
A40
H100
A40
A40
Fine-tuning
time per run
1.0 h
2.9 h
2.1 h
6.2 h
total (3 seeds)
3 h
9 h
6 h
19 h
Deletion test
retraining runs
153
55
153
51
time per run
0.9 h
2.6 h
1.8 h
5.3 h
total
about 140 h
about 145 h
about 275 h
about 270 h
Appendix
Table 10. GPU resources.
Encoder
Checkpoint
Description
CLAP ( Wu et al., 2023 )
laion/larger_clap_music_and_speech
audio encoder trained jointly with a text encoder by contrastive language-audio pretraining
MAEST ( Alonso-Jiménez et al., 2023 )
mtg-upf/discogs-maest-10s-pw-129e
spectrogram-based audio transformer trained by multi-label music style classification
Appendix
Table 11. Audio encoders of the embedding visualizations.
Figure 5. Generated outputs of MusicGen-S in the MAEST embedding space, colored by genre (left) and by the TDA score clusters of GradDot (right).
Figure 6. Generated outputs of MusicGen-L in the MAEST embedding space, colored by genre (left) and by the TDA score clusters of Grad-Cos (right).
Figure 7. Generated outputs of InspireMusic-1.5B in the MAEST embedding space, colored by genre (left) and by the TDA score clusters of TracIn (right).
Figure 8. Generated outputs of MusicGen-L in the CLAP embedding space, colored by genre (left) and by the TDA score clusters of LoGra (right).
Figure 9. Generated outputs of InspireMusic-1.5B in the CLAP embedding space, colored by genre (left) and by the TDA score clusters of FactGraSS (right).