| # Human performance dataset | |
| Piano MIDI dataset used for training and evaluation. Layout and processing: | |
| - **Source:** MAESTRO (or similar) with a `prompt` (or `user_prompt`) per piece. | |
| - **Augmentation:** `dataset_preprocess/augment_dataset.py` adds pitch/time/velocity variants; output CSV lists originals and augmented files. | |
| - **Grouping:** `dataset_preprocess/group_dataset.py` copies MIDIs into `grouped/` as `composer/genre/filename.mid` (and `augmented/composer/genre/` for augmented). Genre is taken from the prompt text. Writes `final.csv` with paths under `grouped/`. | |
| - **Inference CSV:** From `grouped/` (or any `composer/genre/` tree), `src.dataset.generate_prompt_batch` builds `infer.csv` (system_prompt, user_prompt, composer, genre) for batch inference. | |
| See **dataset_preprocess/dataset_preprocess.md** in the repo for commands and details. | |