PearlLeeStudio/TheArtist-MusicTransformer-ft-jazz-only

TheArtist Music Transformer — F5 (Jazz Only, no pop rehearsal)

Jazz-only fine-tune with no pop rehearsal. Reference point for catastrophic forgetting in the companion paper. Strictly dominated by F4 on every axis.

One of six checkpoints released alongside the paper Empirical Study of Pop and Jazz Mix Ratios for Genre-Adaptive Chord Generation (Lee, 2026). See the collection overview at PearlLeeStudio/TheArtist-MusicTransformer-pop-baseline.

Demo

Watch TheArtist in action on YouTube — interactive staff editor, MIDI input, AI generation with live progress, and per-genre LoRA playback across the 13-genre vocabulary.

Model summary

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	7

Training data

All 1,513 jazz training sequences. No pop rehearsal data.

Evaluation (held-out per-genre test sets)

Metric	Pop test	Jazz test
Top-1 accuracy	82.10%	81.30%
Top-5 accuracy	96.31%	92.44%
Perplexity	1.96	2.24
Δ vs. Phase 0 baseline	−2.14	+8.44

F5 illustrates the catastrophic-forgetting failure mode that motivated the paper. Pop accuracy collapses by 2.14 points within a single fine-tune epoch and stabilizes there. Jazz top-1 reaches 81.30%, which is matched by F4 (which also keeps an extra 0.92 points of pop). On every operating axis F5 is dominated by F4, so F5 should not be selected as a production checkpoint. It is released here for replication of the per-epoch forgetting curve and for researchers who want to inspect the failure mode directly.

Known failure modes (this checkpoint specifically)

Chord progressions trend toward dense chromatic voicings that are commercially niche. Generations on pop prompts retain diatonic structure but with persistent chromatic substitution. See paper §6.4 and §7.6 for representative continuations.

Usage

The repo bundles the project's model.py and tokenizer.py at the repo root, so external users can load the checkpoint end-to-end without cloning anything from GitHub. snapshot_download materializes the full repo on disk; sys.path makes the bundled model.py / tokenizer.py importable.

Required dependencies: torch, huggingface_hub.

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-jazz-only")
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 top of this base.

Per-genre real-song eval (held-out 130-song set, 2026-05)

First per-genre evaluation of ft-jazz-only beyond the pop/jazz split that the original paper reports.

Eval results

Genre	n_songs	Top-1 (%)	Top-5 (%)	val_loss
pop	10	85.97	95.78	0.6333
rock	10	85.38	96.46	0.5710
jazz	10	71.16	86.77	1.2883
blues	10	79.84	92.63	0.9230
bossa	10	81.01	95.13	0.8129
classical	10	46.86	77.96	2.3232
country	10	84.81	97.04	0.6185
electronic	10	86.49	97.59	0.5834
folk	10	83.70	98.07	0.6290
funk	10	82.42	94.96	0.7809
gospel	10	78.40	95.58	0.8520
hip_hop	10	89.99	98.12	0.4733
rnb_soul	10	84.20	96.47	0.6757

On this eval set F5 peaks on hip_hop (89.99%) and struggles most on classical (46.86%). This is auxiliary signal — the 11 per-genre LoRAs (sister lora-* repos) are the recommended path for production use on the 9 non-pop, non-jazz genres. F-series cells on those genres show what the base model produces under [GENRE:none] conditioning (the model's [GENRE:X] token does not exist for the 9 new genres in the F-series vocab=351).

Eval dataset composition

130 songs total, 10 per genre × 13 genres. Drawn from the same splits/val.jsonl + splits/test.jsonl partitions every F-series model was held out from during training — no train-set leakage. Built by ai/training/build_eval_real_songs.py --seed 42 --per-genre 10 (deterministic).

Genre	n	Source(s)	Bar range	Avg duration · named
pop	10	billboard	58–116	189s · 10/10 named
rock	10	chordonomicon_rock	52–87	127s · 0/10 named
jazz	10	choco:jazz-corpus, choco:real-book, jazzstandards, jht	16–89	72s · 10/10 named
blues	10	chordonomicon_blues	24–46	93s · 0/10 named
bossa	10	chordonomicon_bossa	24–78	88s · 0/10 named
classical	10	chordonomicon_classical	11–40	60s · 10/10 named
country	10	chordonomicon_country	30–81	110s · 0/10 named
electronic	10	chordonomicon_electronic	25–84	89s · 0/10 named
folk	10	chordonomicon_folk	33–82	114s · 0/10 named
funk	10	chordonomicon_funk	30–60	92s · 0/10 named
gospel	10	chordonomicon_gospel	24–85	98s · 0/10 named
hip_hop	10	chordonomicon_hip_hop	24–81	136s · 0/10 named
rnb_soul	10	chordonomicon_rnb_soul	34–82	128s · 0/10 named

Source license summary: McGill Billboard (CC0, named pop songs), Jazz Harmony Treebank / JazzStandards / WJazzD (Public / community-redistributed, named jazz standards), Bach chorales via music21 (public domain, named pieces), Chordonomicon per-genre subsets (CC BY-NC 4.0; titles are Spotify track IDs by upstream dataset policy — progressions are real songs). See docs/EVAL.md for full breakdown.

Methodology

Teacher-forced next-token cross-entropy / top-1 / top-5 over each song's token sequence (BOS + key + time_sig + genre + bars + EOS, truncated to max_seq_len=256). Same evaluate() call as ai/results/f1_per_genre_baseline.csv, just narrowed to the curated 130-song subset. Token-level metrics; not a generation-quality eval (free-generation comparison with R1 Sethares + R2 theory RAG rerank is documented separately in ai/results/eval_report.md).

Caveats:

• classical val partition is intrinsically small (37 sequences in full eval); the 10-song subset here has even narrower confidence bands. Directional finding (LoRA helps a lot on Bach harmony) is robust, exact pp deltas are noisy.
• F-series numbers on the 9 LoRA-only genres are conditioned without genre tag (vocab=351 has no [GENRE:country] token etc.). This is the realistic "F-series alone" condition, not a controlled ablation.

Source CSV: ai/results/real_song_eval.csv (17 models × 130 songs, long format).

Training-data licenses

Dataset	License
Jazz Harmony Treebank	Public
JazzStandards (iReal Pro)	Community redistribution
Weimar Jazz Database	ODbL
JAAH	Research-use public

Citation

Cite the original mix-ratio paper. The companion per-genre LoRA paper (chord-symbol time-series adaptation) is in preparation; its arXiv ID will be added here once posted.

@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},
  note          = {arXiv preprint, ID TBD},
}