open-rvq-encoder-minimax-music3-169m — 53k base + pooled-corpus fine-tune
Full credit to PurpleOrc for the base model:
PurpleOrc/open-rvq-encoder-minimax-music3-169m-53k,
trained from scratch on their 53k-track (900 h) multilingual self-distilled corpus. This release (v2, 2026-08-21)
fine-tunes their v4_best.pt on the pooled community corpus — 18.7k+ tracks incl. multilingual vocals (Japanese/Spanish/German/Hindi) and a rock/metal/dense-production wave from
Mothersuperior/minimax-music3-rvq-distill-corpus-8k
(diversity-engineered: 355 genres w/ per-genre priors, parity-floor sampling) + the 2,837-track train split of
bghira/minimax-music3-rvq-reverse-distillation —
using PurpleOrc's own recipe (mean CE over 8 heads, teacher-forced depth, bf16) at lr 1e-4 cosine, 8 epochs,
single RTX PRO 6000. Combined, the model has seen ~69k tracks of Music3 self-distillation signal.
Benchmark — bghira's ORIGINAL 130-record exact-alignment holdout
The community-standard holdout: deterministic split from bghira's dataset indexes, never trained on by any encoder in this lineage. Semantic numbers are comparable across all models; acoustic here is TEACHER-FORCED (free-running replay numbers are lower and live in the other model cards).
| model | sem top-1 | sem top-5 | acoustic top-1 (TF) | acoustic top-5 (TF) |
|---|---|---|---|---|
| SimpleTuner v4-base (2,972 tracks) | 0.4323 | 0.8056 | 0.1842* | — |
| Mothersuperior pooled-v5 (14,686) | 0.4745 | 0.8501 | — | — |
| PurpleOrc 53k base | 0.4822 | 0.8467 | 0.2579 | 0.5614 |
| this model v1 (53k + 15.7k pool) | 0.5077 | 0.8781 | 0.2791 | 0.5957 |
| this model v2 (53k + 21.5k pool, multilingual) | 0.5094 | 0.8793 | 0.2800 | 0.5964 |
*v4-base TF acoustic from its model card. First encoder past 50% semantic top-1 on this holdout.
Files / versions
| file | trained on | sem top-1 | note |
|---|---|---|---|
v4_pooled_v1.pt |
53k base + 15.7k pool | 0.5077 | first pooled fine-tune (2026-08-20) |
v4_pooled_v2.pt |
53k base + 21.5k pool (multilingual + heavy) | 0.5094 | current best (2026-08-21) |
v4_pooled_best.pt |
— | — | alias of the current best (= v2); may be overwritten by future versions |
Format / loading
PurpleOrc's architecture (NOT SimpleTuner classes): V4Encoder from the bundled train_v4.py
(1088-wide shared encoder, plain linear readouts, causal depth decoder w/ per-codebook prior embeddings).
import torch
from train_v4 import V4Encoder
model = V4Encoder()
state = torch.load("v4_pooled_best.pt", map_location="cpu", weights_only=False)
model.load_state_dict(state["model"])
Input: 128-frame windows of DAV latents [B, L, 128] + pool matrix [B, 128, L] (stitched-timeline
alignment; see the corpus card's consumption notes). Use is subject to the MiniMax Music 3 model terms.