Mothersuperior/yue2-minted-corpus

YuE2 minted corpus

Songs generated by m-a-p/YuE2-3B (with YuE2-Vae) from MusicForge plan_50k prompts, queued genre-round-robin (355 genres, ~35% instrumental), seeds from the plan. Every track keeps the exact intermediate the model produced, so the set is a labelled corpus for training an audio → semantic-token encoder (the piece YuE2 does not ship) and a grammar regularizer for AR fine-tunes.

tracks/<pid>/ | file | content | |---|---| | audio.flac | 48 kHz stereo 24-bit | | semantic.npy | int32, YuE2 semantic codes (0..32767), 25 per second — the labels | | latent.npy | float32 [T,64] acoustic latents, frame-aligned with semantic.npy | | score.abc, abc_tokens.npy, prefix.npy, plan.json, plan_manifest.json | the AR's score and prompt tokens | | request.json / config.json / result.json | style, lyrics, seed, cot; generation config; hashes + timing | | item.json | queue item (genre, family, bpm, instrumental flag, caption) + wall time |

Regularization pack

regularizer/minted_regularizer_pack.pt — a torch list of {name, src, style, lyrics, codec} for the minted songs (true semantic tokens, no audio; ~60 MB). Mix it 50/50 with an artist's tokenised songs when LoRA-tuning YuE2's AR: it keeps the model's token grammar intact so a small artist set cannot collapse it. src == "minted_val" (5% by pid hash) is a held-out set whose next-token loss should stay flat during training.

import torch; pack = torch.load("regularizer/minted_regularizer_pack.pt", weights_only=False)

Generated with YuE2-3B whose weights are CC BY-NC 4.0; this corpus inherits that license. Growing: uploads are incremental.