PearlLeeStudio/TheArtist-MusicTransformer-ft-pop29

TheArtist Music Transformer — F4 (Pop 1K Mix, jazz-leaning)

Jazz-adapted chord model fine-tuned from the Phase-0 pop baseline with a 1,000-sequence pop rehearsal buffer — the jazz-leaning endpoint of the mix-ratio sweep, one of six checkpoints released with Empirical Study of Pop and Jazz Mix Ratios for Genre-Adaptive Chord Generation (Lee, 2026). It posts the highest jazz top-1 accuracy in the collection (81.50%) at the cost of 1.19 pop points, and is TheArtist app's recommended checkpoint for jazz. Recommended for jazz-flavoured chord composition (chord symbols only — no melody or audio generation); F3 (ft-pop50) is the balanced alternative and F1 (ft-pop80) the pop-leaning endpoint.

Paper · Code · Demo · All models

Served-weights note. The released best.pt is the minimum-mixed-validation-loss checkpoint (epoch 4 — pop 83.38 / jazz 80.68 on the same test sets); the Evaluation table below shows the paper's epoch-6 best-jazz figures. The released F1, cited here as the pop-leaning endpoint, is itself weight-identical to the Phase-0 baseline (a checkpoint-selection artifact — see the F1 card); this checkpoint's weights are hash-distinct from both.

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
Fine-tune resumed from	Phase-0 pop baseline
Best epoch	6 (released best.pt = epoch 4; see note above)

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-ft-pop29")
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 jazz-flavoured continuation.
song = {
    "key": "Cmaj", "time_signature": "4/4", "genre": "jazz",
    "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 — they chain on the released ft-pop80 (F1) base, not this checkpoint.

Evaluation

Held-out per-genre test sets:

Metric	Pop test	Jazz test
Top-1 accuracy	83.02%	81.50%
Top-5 accuracy	96.93%	92.59%
Perplexity	1.80	2.26
Δ vs. Phase-0 baseline	−1.19	+8.64

F4 produces the most jazz-flavoured continuations among the released checkpoints — secondary dominants, tritone substitutions, modal interchange, and ii-V chains across distant keys — at the cost of roughly one point of pop top-1 accuracy.

Per-genre real-song eval

Teacher-forced token-level metrics; on this set F4 peaks on hip_hop (90.34%) and struggles most on classical (49.65%). This is auxiliary signal — the 11 per-genre LoRA adapters are the recommended path beyond pop and jazz — for the eight genres without a matching token in the 351-token vocabulary (all but rock, blues, and bossa) the F-series decodes without genre conditioning.

Genre	n_songs	Top-1 (%)	Top-5 (%)	val_loss
pop	10	85.57	95.91	0.5859
rock	10	87.28	97.50	0.4677
jazz	10	71.42	85.54	1.3367
blues	10	81.99	93.86	0.7970
bossa	10	82.62	95.73	0.7226
classical	10	49.65	81.51	2.1079
country	10	86.30	98.18	0.5191
electronic	10	86.81	98.48	0.5100
folk	10	84.81	98.63	0.5337
funk	10	83.39	96.13	0.6959
gospel	10	80.25	96.58	0.7343
hip_hop	10	90.34	98.58	0.3982
rnb_soul	10	85.04	97.04	0.5907

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).

Training data

All 1,513 jazz training sequences plus 1,000 pop rehearsal sequences (seed 42). Pop:jazz ≈ 0.66:1 — less pop than jazz in the mix. Underlying corpora and licenses: Chordonomicon (CC BY-NC 4.0), McGill Billboard (CC0), Jazz Harmony Treebank (public), JazzStandards / iReal Pro (community redistribution), Weimar Jazz Database (ODbL), JAAH (research-use public).

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}
}