# AI MusiMuse # 011_COMPOSITION_DATASET.md Version 1.0 --- # Purpose The purpose of this document is to define the training data used by the AI Composer. The Composer never trains directly on WAV files. Instead, every analyzed track becomes a structured composition sample. The collection of all composition samples forms the Composition Dataset. --- # Pipeline ``` Audio ↓ Analysis ↓ MusicDNA ↓ Timeline ↓ Sections ↓ Transitions ↓ Composition Graph ↓ Composition Sample ``` --- # Dataset The dataset is simply ``` Collection of Composition Samples ``` Each sample represents one track. --- # Composition Sample Each sample contains ``` Track Metadata MusicDNA Timeline Sections Transitions Energy Curve Density Curve Texture Timeline Motif Graph Style Statistics ``` --- # MusicDNA Global characteristics Examples tempo key spectral profile dynamic profile rhythm profile harmony profile --- # Timeline Ordered musical events Example ``` 0:00 Pad 0:45 Bass 1:12 Theme 2:20 Break 3:00 Theme Returns ``` --- # Sections Example ``` Intro Development Break Outro ``` --- # Transition Objects Example ``` Build Drop Fade Expansion Reduction ``` Every transition contains ``` from to duration energy_delta density_delta confidence ``` --- # Curves Each track contains Energy Curve Density Curve Brightness Curve Motion Curve Harmony Curve These curves are sampled uniformly. --- # Motifs Detected repeating ideas Example ``` Theme A ↓ Variation ↓ Theme A ``` --- # Statistics Every sample stores ``` section count average section duration average transition duration maximum density maximum energy motif count repetition ratio ``` --- # Decision History This is the most important part. The dataset stores ``` State ↓ Decision ↓ New State ``` Example ``` Energy 0.42 ↓ Decision Add Bass ↓ Energy 0.58 ``` --- # Training Record One composition generates many records. Example ``` State ↓ Decision ``` ``` State ↓ Decision ``` ``` State ↓ Decision ``` Thousands of records can be extracted from one song. --- # Dataset Growth The dataset grows automatically. Every analyzed track produces one new sample. Existing samples are immutable. --- # Versioning Every sample stores ``` schema version MusicDNA version analyzer versions builder versions ``` Rebuilding the dataset never overwrites previous versions. --- # Validation Every sample must satisfy deterministic generation no missing sections valid timeline consistent MusicDNA valid transitions non-overlapping events --- # Training Independence The dataset contains no audio. Only musical knowledge. Therefore different renderers do not affect training. --- # Long-Term Goal The Composition Dataset should become a complete representation of the author's compositional language. The AI Composer learns from this dataset, not from raw waveforms.