Downloads · 30 days
0
mrkkucharski/mt3-guitar-pilot
mt3-guitar-pilot is a machine learning model from mrkkucharski. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
T5X/Orbax checkpoints from the MT3 guitar-transcription fine-tuning pilot (see MODALTRAININGSETUPPLAN.md / MT3guitartranscriptionplan.md in the source repo). Not transformers-loadable -- these are raw T5X checkpoint d…
Downloads · 30 days
0
Access
Public
Updated Oct 6, 2026
Repo size
2 GB
Likes
0
Public
Click a slice to open those files.
Other1.5 GB · 90%
From the Hugging Face model README
T5X/Orbax checkpoints from the MT3 guitar-transcription fine-tuning pilot
(see MODAL_TRAINING_SETUP_PLAN.md / MT3_guitar_transcription_plan.md in
the source repo). Not transformers-loadable -- these are raw T5X
checkpoint directories, restorable with this fork's mt3 package.
Decoding settings are not recorded inside a checkpoint. The encoder window, target length and vocabulary flags must match how the checkpoint was trained, or the output degrades silently rather than erroring. Each entry below states them.
297ex_it8_norhythm_pb12_8s_checkpoint_1200000/ — the current working
checkpoint. Trained from the official checkpoint_0 (step 1,000,000) to
step 1,200,000 (200k steps, in four legs) on a 297-example corpus:
48 full-band arrangements plus 249 recorded/rendered guitar examples,
split 238 train / 59 test.
pb12, integer
semitones −12…+12). Decoding this one without pitch bends desyncs the
note decoder, it does not merely drop the bends.Onset + offset + program F1 (full) on its own 59-example test split:
0.8014, the best of the 20 checkpoints evaluated in the run. Treat
that ranking as weak: the final several checkpoints all sit in a
0.78–0.80 band, within run-to-run noise of each other.Kept for comparison. These predate the corpus and vocabulary above, so their metrics are not comparable with each other — each was measured against its own test split.
checkpoint_0/: the official Magenta multi-instrument MT3 checkpoint
(step 1,000,000), unmodified. Re-downloadable from Google's public
bucket; included here for a self-contained repo. 2 s window
(256 frames / 1024 tokens), no rhythm or pitch-bend extensions.replay12s_checkpoint_1290000/: fine-tuned on a mixture of the guitar
corpus and a replay pool of the original MT3 training data, at a 12 s
window (1536 frames / 3072 tokens), rhythm-free, no pitch bends.193ex_it7_norhythm_checkpoint_1196000/: 193-example corpus, rhythm-free,
no pitch bends.70ex_checkpoint_1126000/, checkpoint_1002000/, checkpoint_1040000_it2/,
checkpoint_1078000_4s/, checkpoint_1108000/,
checkpoint_1116020_it3_4s/: earlier engineering and corpus-iteration
legs, superseded. The _4s suffixes denote a 4 s window
(512 frames / 2048 tokens).Any checkpoint whose name contains pb12 must be decoded with pitch bends
enabled; the rest must not be. See PROJECT_LOG.md in the source repo for
per-run training curves, corpus composition and caveats.