open-rvq-encoder-minimax-music3-169m — pooled-corpus fine-tune (pooled-v4)
SimpleTuner/open-rvq-encoder-minimax-music3-169m-v4
(169M, causal depth decoder) fine-tuned from its final checkpoint on a pooled corpus 3.6x its original data:
the original 2,837 train tracks + 8,122 tracks from
Mothersuperior/minimax-music3-rvq-distill-corpus-8k.
Architecture, loss (CE + 0.25 KL vs teacher top-50), muP config, and trainer unchanged — only the data.
12 epochs, lr 3e-4 polynomial, batch 64, single RTX PRO 6000, validated every 500 steps on the untouched
135-track holdout.
Benchmark (same harness/holdout; running the base through it reproduces its published 0.8748 / 0.432 exactly)
| metric | v4 base | pooled-v4 | 41m-pooled-v3 |
|---|---|---|---|
| replay conditioning cosine (mean) | 0.8748 | 0.8936 | 0.7843 |
| semantic top-1 | 0.4323 | 0.4701 | 0.4556 |
| semantic top-5 | 0.8056 | 0.8449 | 0.8322 |
| acoustic top-1 | 0.0730 | 0.0862 | 0.0853 |
| acoustic top-5 | 0.1992 | 0.2243 | 0.2407 |
Raw eval outputs in evaluation/. Loading identical to the base model (trainer classes from SimpleTuner
branch script/train-minimax-music-rvq-encoder, --depth_decoder topology).
Use is subject to the MiniMax Music 3 model terms and the reverse-distillation dataset terms.