Notes You Can’t Delete
depth
home/Models & data
Models & data

Download the models

Every checkpoint is a small from-scratch transformer — 25 to 26 million parameters, about 100 MB each — on a public HuggingFace repo, with the code to run them.

Run one in three lines

pip install torch numpy miditok symusic huggingface_hub
# model.py + generate_example.py are in the repo
python generate_example.py --model g_metal_s1   # writes out.mid

Each checkpoint is a state dict plus its config; the matching tokenizer ships alongside. The whole point is the pairs — a control and its edited twin, differing in exactly one thing — so you can probe the reconstruction yourself.

The checkpoints

modelwhat it is
g_metal_band_s1Full bandguitar + bass + drums, ctx 2048, 316M tokens · val 0.18
g_metal_s1Guitar base22,037 metal tracks, 163M tokens · val 0.23
g_metal_deb5The b5-deleted twinnever saw a tritone — still expects it at 99%
g_metal_td_s2 / g_metal_bm_s3Subgenre fine-tunestech-death and black metal, 12-second edits
g_metal_minus_bm / dj / randLineage cutsa genre (or random data) removed
blues_control_s0 / deblue_s0Blue-note deletionthe matched jazz pair
inject_control / loose / cleanBlue-note injectionclassical, two doses
remap_permuted_s0Scrambled tonal functionun-scrambles to coherent tonality
+ 8 morethe full setinvented rule, chord-melody, dom7 pilot, drums-only…

The training data

All public: the Weimar Jazz Database (jazz solos with ground-truth chords), MAESTRO v3 (classical piano), and a Kaggle metal-MIDI dump (PDDL). Prefer to just hear them? Every finding has playable audio, and the playground runs the tiny models live in your browser.