| --- |
| license: mit |
| language: |
| - ru |
| library_name: onnxruntime |
| pipeline_tag: token-classification |
| tags: |
| - punctuation |
| - capitalization |
| - russian |
| - onnx |
| - int8 |
| --- |
| |
| # scribe-punct-ru |
|
|
| Russian punctuation + capitalization models for [Scribe](https://github.com/AIgrator/Scribe) |
| offline voice dictation, exported to ONNX and quantized to int8 for CPU inference. |
|
|
| The repo hosts two model sizes. Scribe downloads the one listed in its |
| [`punct_models.json`](https://github.com/AIgrator/Scribe/blob/main/punct_models.json) |
| catalog, into `models/punct/ru/`. |
|
|
| ## Models |
|
|
| | Model | Files | Size | Speed (CPU) | Notes | |
| |---|---|---|---|---| |
| | **RUPunct_big** (BERT-base) | `rupunct_big_int8.onnx`, `rupunct_big_tokenizer.json`, `rupunct_big_model.json` | ~178 MB | ~15 ms per phrase | Best quality; default in the Scribe catalog | |
| | RUPunct_small | `rupunct_small_int8.onnx`, `tokenizer.json`, `model.json` | ~29 MB | ~1–3 ms per phrase | Lightweight alternative, kept for compatibility | |
| |
| Both models share the same 33-label scheme (case of *this* word + |
| mark after it), so the same decoding code in Scribe works for both. |
| The sidecar `*_model.json` is the Scribe label map and decoding style description. |
| |
| ## Label scheme |
| |
| 33 classes: case of **this** word (`UPPER` / `LOWER` / `TOTAL`) + mark after it |
| (`PERIOD`, `COMMA`, `QUESTION`, `TIRE`, `VOSKL`, `DVOETOCHIE`, `PERIODCOMMA`, |
| `DEFIS`, `QUESTIONVOSKL`, `MNOGOTOCHIE`, `O`). |
| |
| ## Usage in Scribe |
| |
| Downloaded by Scribe's model manager into `models/punct/ru/`, runs fully offline |
| via ONNX Runtime. No `torch` / `transformers` needed at runtime. |
| |
| ## License / attribution |
| |
| MIT. |
| |
| - **RUPunct_big**: base model [`RUPunct/RUPunct_big`](https://huggingface.co/RUPunct/RUPunct_big); |
| fp32 ONNX export by [`IcPlus/rupunct-onnx`](https://huggingface.co/IcPlus/rupunct-onnx) (MIT); |
| dynamic int8 quantization is a format conversion only, weights were not retrained. |
| - **RUPunct_small**: export of `ekhodzitsky/rupunct-small-onnx` (MIT), converted |
| from `RUPunct/RUPunct_small`. |
| |
| Scribe sidecars (`model.json`, `rupunct_big_model.json`) and this card are MIT as well. |
| |
