Add model card
Browse files
README.md
ADDED
|
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
base_model: OpenMOSS-Team/MOSS-Transcribe-preview-2B
|
| 4 |
+
library_name: mlx
|
| 5 |
+
tags:
|
| 6 |
+
- mlx
|
| 7 |
+
- asr
|
| 8 |
+
- speech-to-text
|
| 9 |
+
- audio
|
| 10 |
+
- moss
|
| 11 |
+
pipeline_tag: automatic-speech-recognition
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
# MOSS-Transcribe-preview-2B — MLX conversion
|
| 15 |
+
|
| 16 |
+
MLX-weight conversion of [`OpenMOSS-Team/MOSS-Transcribe-preview-2B`](https://huggingface.co/OpenMOSS-Team/MOSS-Transcribe-preview-2B), an English speech-to-text model pairing a Qwen3-1.7B-base decoder with a Qwen3-Omni-MoE audio encoder and a gated-MLP adapter (~2.4B parameters, Apache-2.0 upstream).
|
| 17 |
+
|
| 18 |
+
Converted from the upstream bf16 safetensors by the [peacock-asr `moss-mlx-conversion` project](https://github.com/chikingsley/peacock-asr/tree/main/projects/moss-mlx-conversion).
|
| 19 |
+
|
| 20 |
+
## Variants
|
| 21 |
+
|
| 22 |
+
| Directory | Precision | Size | Notes |
|
| 23 |
+
| --- | --- | --- | --- |
|
| 24 |
+
| `MOSS-Transcribe-preview-2B-bf16/` | bf16 | 4.6 GB | Reference MLX-layout conversion; matched PyTorch transcript exactly on the LibriSpeech smoke fixture. |
|
| 25 |
+
| `MOSS-Transcribe-preview-2B-text-decoder-4bit-g64/` | 4-bit affine, group 64 (text decoder only) | 2.7 GB | Strongest quantized candidate in the conversion project: 2.48 RTFx on the first 20 LibriSpeech clean-test rows with no bf16 WER regression on that slice. |
|
| 26 |
+
| `MOSS-Transcribe-preview-2B-all-4bit-g64/` | 4-bit affine, group 64 (all weights) | 1.3 GB | Full-model 4-bit; retained as a size-optimal candidate. |
|
| 27 |
+
|
| 28 |
+
Each variant ships `config.json`, the Qwen ByteLevel tokenizer files, and the conversion/quantization reports produced by the converter.
|
| 29 |
+
|
| 30 |
+
## Validation (Apple Silicon, conversion project)
|
| 31 |
+
|
| 32 |
+
- LibriSpeech clean-test 100-row baseline: **1.80% WER**, 1.61 RTFx (MLX bf16), vs 2.01% WER on the PyTorch bf16 reference.
|
| 33 |
+
- Weight-load, fixture-transcription, and streamed-eval gates passed on Apple Silicon.
|
| 34 |
+
|
| 35 |
+
## Usage
|
| 36 |
+
|
| 37 |
+
**These weights are not loadable by vanilla `mlx-lm`.** MOSS-Transcribe is a multimodal audio-encoder + adapter + Qwen3-decoder architecture with no upstream mlx-lm model type. Load them with the conversion project's runtime:
|
| 38 |
+
|
| 39 |
+
```bash
|
| 40 |
+
git clone https://github.com/chikingsley/peacock-asr
|
| 41 |
+
cd peacock-asr/projects/moss-mlx-conversion
|
| 42 |
+
uv run moss-streaming-eval --weights <variant-dir> # streamed LibriSpeech WER eval
|
| 43 |
+
```
|
| 44 |
+
|
| 45 |
+
See the project `README.md` and `docs/PROGRESS.md` for the smoke test (`moss-mlx-smoke`), the local backend (`MossTranscribeBackend.generate`), and eval harness details.
|
| 46 |
+
|
| 47 |
+
## License
|
| 48 |
+
|
| 49 |
+
Apache-2.0, matching the upstream model. If you use these weights, also cite the upstream model:
|
| 50 |
+
|
| 51 |
+
```bibtex
|
| 52 |
+
@misc{moss_transcribe_2025,
|
| 53 |
+
title = {{MOSS-Transcribe-preview-2B}},
|
| 54 |
+
author = {{OpenMOSS Team}},
|
| 55 |
+
year = {2025},
|
| 56 |
+
howpublished = {\url{https://huggingface.co/OpenMOSS-Team/MOSS-Transcribe-preview-2B}}
|
| 57 |
+
}
|
| 58 |
+
```
|