Mothersuperior/minimax-music3-rvq-distill-corpus-8k

MiniMax Music3 self-distillation corpus — 11,847 tracks / ~193 h

Paired (audio, RVQ codes, teacher top-50 distributions, DAV latents) traces generated with the official MiniMax Music3 pipeline (diffusers ModularPipeline), built to train/improve the community audio->codes RVQ encoder. Fine-tuning SimpleTuner/open-rvq-encoder-minimax-music3-41m-v1 on this corpus pooled with its original data improves every holdout metric — see Mothersuperior/open-rvq-encoder-minimax-music3-41m-pooled-v3.

Diversity engineering: captions come from a genre-first coherence engine — 355 genres/subgenres each with BPM band, meter distribution, key-quality weights, instrumental probability, ensemble templates, structure and duration priors — drawing on ~1,760 curated descriptors across 17 banks (274 instruments, era-matched production 1920s-2020s, moods, contexts, tempo-feel, 16 meters w/ odd meters overweighted, 131 keys/modes, 65 structures). Parity-with-floors sampling guarantees every genre family and meter a minimum share; the corpus overweights rock/metal/punk and dense-production material relative to the plan's parity targets (each track's plan metadata allows re-balancing at training time). Durations 30-150 s (mean ~58.5 s realized). All lyrics are original LLM-written sheets matched to genre diction (English-only; no copyrighted lyrics), embedded in each track's JSON.

KNOWN CONSUMPTION PITFALLS — read before training (credit: scragnog)

Two bugs bite naive consumers of this corpus; both fixes are lossless. Documented in full in scragnog/open-rvq-encoder-minimax-music3-169m-hotstep-v1.

1. Semantic teacher ids are uint16-wrapped LM token ids. probs.npz idx[:, 0, :] (head 0 only) holds raw language-model token ids (code + 151675) stored in uint16, which wraps them mod 65536. Recover the true semantic codes with sem = idx[:, 0, :].astype(np.int32) - 20603, then mask entries outside [0, 16384) (e.g. the audio-end token). Consumed naively, the semantic ids are garbage and a teacher-KL/soft-CE loss on head 0 is meaningless (or, with range-masking trainers like SimpleTuner's, silently skipped). Heads 1-7 are unaffected (plain 0-1023). Verified: after unwrapping, 100% of semantic entries are valid and teacher-top1 matches the sampled code at the expected rate.

2. Use the stitched-timeline frame->latent mapping, not uniform. Codes and vae.npy latents align via the 200-frame DiT window / 100-frame hop / 345-latent stitched timeline (legacy_nominal in SimpleTuner's trainer). A naive uniform total_latents / total_frames mapping drifts ~1 latent per 800 frames — semantic accuracy decays to noise past the first minute of a track. scragnog found a constant +1 latent offset empirically best: k = 0 if i < 125 else (i - 25) // 100; start = 345 * k + ((i - 100 * k) * 441) // 128 + 1. Real audio needs no such correction.

Per track (zip shards under data/, ~250 tracks each)

file	contents
<id>.flac	44.1 kHz stereo audio (lossless from the pipeline output)
<id>.codes.npy	int32 [frames, 8] sampled RVQ codes; row 0 is the un-emitted priming row (emitted frame i = codes[i+1])
<id>.probs.npz	teacher top-50 per head from the FINAL guided logits at each sampling call: idx uint16 [frames, 8, 50], prob float16 (renormalized softmax; log(prob) reproduces top-k logits for KL distillation)
<id>.vae.npy	float16 [T, 128] DAV latents of the rendered audio (official dav.pth encoder)
<id>.json	caption, lyrics, seed, duration, timing, alignment validation

Alignment: constant stitched timeline (200-frame DiT windows / 100-frame hop, 345-latent hop, non-first-chunk ownership from frame 25, per-chunk floor(F*441/128)) — identical to the legacy_nominal rule in SimpleTuner's train_minimax_music_rvq_encoder.py; every track validated at generation (11,847/11,847 align_ok and probs_ok; see manifest). manifest.jsonl lists every track; corpus_plan/ has the full 50k prompt plan + coverage ledger (this release is the 11,847 minted so far from that plan).

All audio is Music3 self-generated from original prompts/lyrics; use is subject to the MiniMax Music 3 model terms.