HaileyStorm/sv-billie-yue2-lora

sv-billie · YuE2 LoRA

A YuE2-3B adapter for close, hushed vocals, heavy low end and off-kilter bedroom pop... sounds... familiar...

Selected checkpoint: step 950. Rank-32 PEFT-format adapter trained with AI Toolkit on an artist-focused recorded-music corpus. Training audio is not redistributed. This repository contains the adapter, not the base model or VAE.

Listen: The Long Way Back album

Eight original generated songs. 3 seeds x 3 sampler configs, best generation picked (some songs all generations discarded)- no edits or other manual work. Except for Porch Light, which was one generation per LoRA checkpoint.

For best demonstration of LoRA result, start with Smile for the Meter, then The Fridge Light and No Applause in the Kitchen. Album files: MP3, FLAC, WAV, lyrics and style prompts.

01. Smile for the Meter · MP3

02. Velvet on the Handbrake · MP3

03. The Seat Beside You · MP3

04. Spare Key · MP3

05. Porch Light · MP3

06. The Fridge Light · MP3

07. No Applause in the Kitchen · MP3

08. Last Ferry, First Bird · MP3

Porch Light is an earlier step-900 example. Last Ferry, First Bird uses step 950 plus a DreamPop adapter at 0.3, which is not included here. Other album tracks use this artist adapter alone. Some endings await manual trimming. Exact selected settings and seeds are in the album manifest. These are curated examples, not a guarantee of consistent quality or lyric adherence.

Optional support

This adapter is free and donations are entirely optional. Donations help sustain Hailey's hands-on AI work: past chess and Mamba-model work, and current or future work around Maestro Continuum, Palimpsest, Loom/ScaFOLD, an experimental Qwen3.8-27B abliteration/efficient-reasoning model, other tools, and whatever comes next. Donations never privilege or discourage any model, provider, or host route, and no available donation method below is preferred:

• Bitcoin: 1K628QLEh3sS8sEdzZfvuqqHRecVckSgaJ
• Cardano: addr1q9fd05jktgv49094z8hvjp6cqvn7npt8hfzjna4dvhezmvpgl92x5cevqghl4ng0we2es4xjp59gvm3nttdzwf9ym6lqr3628x
• Ethereum: 0x78b6adac22415568A7F725a865206ccFd1a82F4c
• Vast.ai credit: transfer credit to HaileyCollet@gmail.com. Vast.ai says transfers are irreversible, so double-check the recipient before confirming.
• Buy Me a Coffee: one-time support and memberships through the verified public Threadspan page.

See the donation instructions for QR codes, Vast.ai instructions, and safety notes.

Use

Load sv_billie.safetensors with a YuE2-compatible LoRA loader (AR + NAR), starting at strength 1.0. Use YuE2-3B and YuE2-Vae. The included lora.json provides the display name and trigger for Sound and Vision.

Prepend sv_billie, in the style of sv_billie. to your style caption, then specify genre, instruments, language and tempo. Supply full section-tagged lyrics. Album ABC plans and lyric files are separate; supplied ABC bypasses score planning.

Profile	AR temperature	top-p	top-k	repetition penalty
Default	0.88	0.96	115	1.20
Focused	0.82	0.94	100	1.24
Exploratory	0.94	0.975	130	1.18

Use 100 ODE steps, CFG 1; score sampling 0.8 / 0.925 / 40 when planning. These are general starting profiles; album takes used individual variants. Choose generous duration headroom. Random seeds are recommended for exploration; 2029 is a convenient fixed seed, not a proven quality improvement.

Tokenizer and training scripts

See https://huggingface.co/Mothersuperior/yue2-mothersuperior-realaudio-tokenizer-v4.

License and attribution

Adapter: CC BY-NC 4.0, following the non-commercial base-model license. Credit the YuE2 team and this adapter repository. The base license does not grant rights to third-party recordings, identities or likenesses. Samples are AI-generated demonstrations with newly written lyrics; no source recordings are included.