The model is designed for cropped single-system guitar tablature images, especially digital score exports from Guitar Pro, alphaTab, MuseScore-style renderers, and similar notation tools. It predicts a token sequence describing bars, beats, durations, notes, rests, tuplets, tuning/capo metadata, and common guitar techniques.

It is not intended as a general OCR model, a full score editor, or a reliable parser for arbitrary scanned sheet music.

Files

best.pt: best validation checkpoint.
config.json: training and model configuration.
vocab.json: OMR token vocabulary.
summary.json: best checkpoint summary.
metrics.jsonl: training history.
eval-greedy.json/metrics.json: greedy decode evaluation metrics.
eval-constrained-beam.json/metrics.json: constrained beam evaluation metrics.

Model

Architecture: ConvNeXt tiny encoder + branch transformer decoder.
Input size: 1200 x 224, grayscale.
Max decode length: 384.
Hidden size: 256.
Decoder layers: 3.
Attention heads: 4.
Feed-forward size: 1024.

Evaluation

Validation was run on the renderer-augmented Tabbox guitar OMR dataset.

Decode	Samples	Exact Match	Parse Success	Core SER	Token Edit Similarity
Greedy	200	76.5%	99.5%	0.0236	97.64%
Constrained beam	200	76.5%	99.5%	0.0240	97.60%

Best checkpoint:

Best metric: val_core_ser
Best value: 0.00852
Best epoch: 24

Limitations

Works best on clearly cropped guitar tab systems, not full PDF pages.
Training data is still biased toward clean digital exports.
Some uncommon notation layouts, handwritten scans, dense multi-voice notation, and unusual renderer styles may degrade token quality.
Predicted tokens may require downstream validation before converting to a playable score.

Usage

This model uses custom model and decoding code rather than a stock Transformers pipeline. For a simple interactive trial, use the Space linked above. For programmatic use, load best.pt, config.json, and vocab.json with the inference code in the Tabbox repository.