PearlLeeStudio/TheArtist-MusicTransformer-ft-pop80-v2

TheArtist Music Transformer — F1 v2 (Pop 10K Mix, selection-corrected retrain)

Selection-corrected retrain of the F1 slot: the Phase-0 pop baseline fine-tuned on jazz with a 10,000-sequence pop rehearsal buffer, identical in recipe to ft-pop80 except that the best checkpoint is selected on a jazz-only validation subset (select_best_on: jazz_val) instead of the pop-dominated validation mix. Held-out jazz top-1 reaches 80.31 (+7.45 pp over the un-adapted state) at a −0.48 pp pop cost. One of the models from the paper How Far Can Chord-Symbol Time-Series Adaptation Carry Genre Identity? (Lee, 2026).

Recommended for chord-composition workflows targeting pop/rock with genuine jazz coloration — jazz substitutions (secondary dominants, ii–V detours) appear selectively over a pop harmonic root. Choose ft-pop29 (F4) when jazz is the primary target.

Paper · Code · Demo · All models

Why a v2 exists. The v1 release (ft-pop80) shipped best.pt weights that coincide with the Phase-0 pop baseline — a checkpoint-selection artifact explained in the note on that card. This retrain changes only the selection metric; its best.pt is verified hash-distinct from Phase-0 (full-state SHA-256 b47144d4db058638… vs 3875f40784f19811…) and embodies the jazz-leaning adaptation the F1 slot was designed for. The 11 per-genre lora-* adapters were trained on the v1 released base and remain paired with it, not with this checkpoint — use this repo as a standalone jazz-leaning base.

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 (selected by jazz-only val loss, 0.9331)
Trained on	Google Colab Tesla T4, AMP, ~20 min/epoch

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-pop80-v2")
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()))

Evaluation

Teacher-forced metrics on the held-out per-genre test sets, best epoch 6:

Metric	Pop test	Jazz test (6-src, 167 sequences)
Top-1 accuracy	83.73%	80.31%
Top-5 accuracy	96.95%	92.42%
Perplexity	1.78	2.31
Δ vs. un-adapted init	−0.48	+7.45

The headline jazz numbers are measured on the 6-source jazz test (167 sequences); on the unified 9-source jazz test (501 sequences) the same weights score 75.05, so figures across the two test sets are not cell-comparable. For this run (seed 42, 6-source corpus) the mixed validation loss rose monotonically (0.5578 → 0.5929) — min-mixed-val selection would have retained the un-adapted epoch 3, while jazz-only selection picks epoch 6. In matched-data multi-seed retrains on the 9-source corpus (seeds 42, 123, 7), jazz-val selection reliably yields an adapted base: jazz top-1 75.76 ± 0.03, all hash-distinct from Phase-0. The full per-epoch training curve, including the jazz_val_* selection-metric columns, ships in this repo's eval_results.csv.

Per-genre real-song eval

Paired against Phase-0 on the same songs, this checkpoint gains +5.17 pp mean jazz top-1 (median +5.19, 6/10 songs win) at −0.42 pp pop, with all other genres within ±1.1 pp. For the eight genres without a [GENRE:X] token in the 351-token vocabulary (all beyond pop/jazz except rock, blues, and bossa) the model runs without a genre tag — the per-genre lora-* adapters, paired with the v1 base, are the recommended path there. 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).

Genre	n_songs	Top-1 (%)	Top-5 (%)	val_loss
pop	10	86.26	95.74	0.6167
rock	10	86.59	97.32	0.4877
jazz	10	70.13	86.42	1.3382
blues	10	82.61	94.26	0.7972
bossa	10	82.02	96.00	0.7641
classical	10	49.27	81.63	2.1650
country	10	85.62	98.04	0.5327
electronic	10	86.85	98.50	0.5087
folk	10	84.58	98.62	0.5525
funk	10	83.12	95.79	0.7053
gospel	10	80.10	96.76	0.7609
hip_hop	10	90.09	98.40	0.3965
rnb_soul	10	84.95	96.74	0.6152

Training data

All 1,513 jazz training sequences (Jazz Harmony Treebank, JazzStandards, Weimar Jazz Database, JAAH) plus 10,000 pop rehearsal sequences sub-sampled with seed 42 from the Phase-0 pop training split — the same mix as v1, drawn from the rebuilt 6-source splits.

Fine-tune hyperparameters: peak learning rate 2 × 10⁻⁵, two-epoch warmup, ten epochs maximum with patience 5 — identical to v1; only select_best_on: jazz_val differs.

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}
}