Aria-medium-GiantMIDI

loubb/aria-medium-base (an autoregressive solo-piano MIDI model) adapted to GiantMIDI-Piano with LoRA (r=16), merged into the full weights. This is the "in-domain pretraining" reference arm of a CSE 253 (UCSD) study.

Usage

Load with trust_remote_code=True. The custom modeling code is bundled. Use the tokenizer from the base model loubb/aria-medium-base (encode_from_file, decode().to_midi()).

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("sullivanUCSD/Aria-medium-GiantMIDI", trust_remote_code=True)
tok = AutoTokenizer.from_pretrained("loubb/aria-medium-base", trust_remote_code=True)

Results (test set, own 17,727-token vocabulary)

Perplexity 8.05, interval KL 0.093, repetition 0.0002, 0% invalid.

Base model and tokenizer: Aria (Bradshaw & Colton, ICLR 2025). Please follow the base model's license. Coursework artifact.