README.md

---
license: apache-2.0
library_name: safetensors
tags:
  - music
  - lyrics2song
  - yue
  - sceneworks
---

# xcodec-mini-infer — SceneWorks safetensors rehost

A safetensors-only redistribution mirror of [`m-a-p/xcodec_mini_infer`](https://huggingface.co/m-a-p/xcodec_mini_infer) at revision
`fe781a67815ab47b4a3a5fce1e8d0a692da7e4e5` — the xcodec codec (SoundStream/SEANet + RVQ + HuBERT semantic branch), the two Vocos
44.1 kHz upsamplers (vocal `decoder_131000`, instrumental `decoder_151000`) and the mm tokenizer
used by YuE — published by **SceneWorks** for SceneWorks Inference's candle YuE engine. It is
**not** an official M-A-P distribution.

## Contents

| File | Tensors | dtype | sha256 |
|---|---|---|---|
| `final_ckpt/ckpt_00360000.safetensors` | 530 | float32 | `b8467e2173e15ffa…` |
| `decoders/decoder_131000.safetensors` | 81 | float32 | `de479526121d6be8…` |
| `decoders/decoder_151000.safetensors` | 81 | float32 | `e961e2c9f1220bdc…` |
| `semantic_ckpts/hf_1_325000/model.safetensors` | 211 | float32 | `2ac43d6c891ac8cc…` |

Each converted file is tensor-equal (value, dtype, shape) to the upstream pickle. From
`ckpt_00360000.pth` only the `codec_model` state dict is kept; the training-only optimizer,
LR-scheduler and discriminator (`mfd`) entries are dropped. The YAML configs, the HuBERT
`config.json` / `preprocessor_config.json`, and `mm_tokenizer_v0.2_hf/tokenizer.model` are copied
verbatim; `mm_tokenizer_v0.2_hf/tokenizer.json` is derived (see the LM repos' READMEs).

No Python code from the upstream repo is redistributed.

## Tiers

Not tiered — approved carve-out (epic sc-19373 R2): every LM tier (bf16 / q8 / q4) uses these same
files at upstream precision. `sceneworks-tiers.json` records this.

## License

**Apache-2.0**, © 2025 Ruibin Yuan and core contributors from M-A-P and HKUST. See `LICENSE` and
`NOTICE` (retained per Section 4(d) of the Apache License).