umutgur/t2p

BachGround Text to Piano

This Hugging Face repository contains the public model artifacts for the BachGround text-to-piano pipeline.

The pipeline has two stages:

1. a Llama 3.1 based adapter that generates base symbolic piano tokens from text
2. a complementary transformer that predicts duration and velocity tokens to enrich the base sequence before MIDI rendering

Repository Layout

• models/t2p_base Llama-based adapter files and tokenizer assets
• models/complementary_transformer Complementary transformer code and trained duration/velocity models
• scripts/infer_music_full_pipeline.py End-to-end prompt -> base tokens -> enrichment -> MIDI pipeline
• scripts/infer_llama_music.py Optional standalone base-model inference
• scripts/midi_to_mp3.py Optional MIDI to MP3 rendering utility

Usage

Use the end-to-end pipeline script for the normal inference path:

python3 scripts/infer_music_full_pipeline.py \
  --llama-model-dir models/t2p_base \
  --duration-model-dir models/complementary_transformer/models/duration \
  --velocity-model-dir models/complementary_transformer/models/velocity \
  --prompt "A calm piano melody in C major" \
  --output-dir pipeline_out \
  --do-sample \
  --temperature 0.9 \
  --top-p 0.9 \
  --max-new-tokens 450

License Breakdown

This repository uses the llama3.1 Hugging Face license tag because it contains a Llama 3.1-derived adapter.

Artifact-level intent:

• models/t2p_base/* Subject to the applicable Llama 3.1 license terms
• models/complementary_transformer/* Intended to be released under separate terms; see the component README before public release

Links

• Project GitHub page: https://github.com/BachGround/t2p
• Huggingface repo: https://huggingface.co/umutgur/t2p
• Website: https://www.bachground.com

Notes

• This repository is designed for inference and artifact distribution.
• Training datasets are not included.
• If you later decide to separate the weights into multiple Hugging Face repositories, keep this README as the umbrella project card and link out to the split repos.