PearlLeeStudio/TheArtist-MusicTransformer-pop-baseline

TheArtist Music Transformer — Phase-0 (Pop Baseline)

Pop-only pretraining baseline of the mix-ratio study — a Music Transformer trained from scratch on ~544K pop songs × 3 epochs (Chordonomicon + McGill Billboard). All five F-series jazz fine-tunes resume from this checkpoint, making it the no-jazz reference for measuring catastrophic forgetting in the paper How Far Can Chord-Symbol Time-Series Adaptation Carry Genre Identity? (Lee, 2026).

It is the most pop-fluent and most jazz-naive model in the collection — use it when pop output is the only target; for balanced pop/jazz output use F3 (ft-pop50), for jazz-leaning output F4 (ft-pop29).

Paper · Code · Demo · All models

Model details

Field	Value
Architecture	Music Transformer with relative positional attention
Parameters	25,661,440
Vocabulary size	351 tokens
Max sequence length	256
d_model / heads / FFN / layers	512 / 8 / 2048 / 8
Training framework	PyTorch 2.5+ (CUDA 12.1)

Usage

Requires torch, huggingface_hub. The repo bundles model.py and tokenizer.py, so nothing needs to be cloned from GitHub.

import sys
import torch
from huggingface_hub import snapshot_download

# Download the full repo (model.py, tokenizer.py, best.pt, config.json).
ckpt_dir = snapshot_download(repo_id="PearlLeeStudio/TheArtist-MusicTransformer-pop-baseline")
sys.path.insert(0, ckpt_dir)  # so the next two imports resolve

from model import MusicTransformer
from tokenizer import ChordTokenizer

tokenizer = ChordTokenizer()
ckpt = torch.load(f"{ckpt_dir}/best.pt", map_location="cpu", weights_only=False)
model = MusicTransformer(
    vocab_size=tokenizer.vocab_size,
    d_model=512, n_heads=8, d_ff=2048, n_layers=8,
    max_seq_len=256, dropout=0.0, pad_id=tokenizer.pad_id,
)
model.load_state_dict(ckpt["model_state_dict"])
model.eval()

# Prompt = ii-V-I in C major; ask for a pop-flavoured continuation.
song = {
    "key": "Cmaj", "time_signature": "4/4", "genre": "pop",
    "bars": [["Dm7", "G7"], ["Cmaj7"]],
}
prompt_ids = tokenizer.encode_sequence(song)[:-1]
ids = torch.tensor([prompt_ids])
with torch.no_grad():
    for _ in range(32):
        logits = model(ids)
        next_id = torch.multinomial(
            torch.softmax(logits[:, -1, :] / 0.8, dim=-1), 1,
        )
        ids = torch.cat([ids, next_id], dim=-1)
        if next_id.item() == tokenizer.eos_id:
            break
print(tokenizer.decode(ids[0].tolist()))

For per-genre adaptation beyond pop and jazz, see the 11 LoRA adapter repos at PearlLeeStudio.

Evaluation

Held-out per-genre test sets:

Metric	Pop test	Jazz test
Top-1 accuracy	84.21%	72.86%
Top-5 accuracy	97.09%	86.51%
Perplexity	1.73	4.01

The 72.86% jazz top-1 from a pop-only model reflects the substantial token overlap between the two genres; jazz-specific gains in the fine-tuned checkpoints come from learning the transition statistics over those shared tokens, not from learning new tokens.

Per-genre real-song eval

Teacher-forced next-token top-1 / top-5 / cross-entropy over each song's full token sequence (truncated to max_seq_len=256). On this set Phase-0 peaks on hip_hop (90.66% top-1) and struggles most on classical (49.55%).

Genre	n_songs	Top-1 (%)	Top-5 (%)	val_loss
pop	10	86.68	96.01	0.5734
rock	10	86.69	97.48	0.4578
jazz	10	64.96	81.16	1.8958
blues	10	81.52	93.91	0.8410
bossa	10	81.43	95.47	0.7825
classical	10	49.55	81.17	2.2389
country	10	85.90	98.44	0.5152
electronic	10	87.39	98.45	0.5072
folk	10	85.04	98.92	0.5244
funk	10	83.85	96.03	0.6811
gospel	10	79.79	96.85	0.7367
hip_hop	10	90.66	98.59	0.3957
rnb_soul	10	85.10	97.07	0.5877

130 songs (10 per genre × 13 genres, seed 42) drawn from held-out val/test partitions — pop from McGill Billboard (CC0), jazz from public standards corpora, classical from Bach chorales, the other ten genres from the matching Chordonomicon subsets (CC BY-NC 4.0; titles are Spotify track IDs by upstream policy). Source license summary: McGill Billboard (CC0), Jazz Harmony Treebank / JazzStandards / WJazzD (Public), Bach chorales (Public Domain), Chordonomicon per-genre subsets (CC BY-NC 4.0).

Training data

Trained from scratch on the pop training split — ~544K songs drawn from the Chordonomicon dataset (CC BY-NC 4.0) and McGill Billboard (CC0), lightly filtered and twelve-key-augmented. Three epochs at peak learning rate 3 × 10⁻⁴ with one-epoch warmup and cosine decay; wall-clock time ≈27 hours on a single NVIDIA GeForce RTX 4070 Laptop GPU.

License

CC BY-NC 4.0 (weights; matching Chordonomicon, the dominant training corpus). Research, paper replication, portfolio, and demo use are permitted; commercial use is not.

Citation

@misc{lee2026chordmix,
  title         = {Empirical Study of Pop and Jazz Mix Ratios for Genre-Adaptive Chord Generation},
  author        = {Lee, Jinju},
  year          = {2026},
  eprint        = {2605.04998},
  archivePrefix = {arXiv}
}

@misc{lee2026chordtimeseries,
  title         = {How Far Can Chord-Symbol Time-Series Adaptation Carry Genre Identity?},
  author        = {Lee, Jinju},
  year          = {2026},
  eprint        = {2606.07334},
  archivePrefix = {arXiv}
}