Token Classification
Transformers
ONNX
Safetensors
Danish
electra
punctuation-restoration
sentence-boundary-detection
truecasing
danish
Instructions to use RyeAI/ekko-pnc with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use RyeAI/ekko-pnc with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("token-classification", model="RyeAI/ekko-pnc")# pip install -U transformers accelerate # Load model directly from transformers import AutoTokenizer, AutoModelForTokenClassification tokenizer = AutoTokenizer.from_pretrained("RyeAI/ekko-pnc") model = AutoModelForTokenClassification.from_pretrained("RyeAI/ekko-pnc", device_map="auto") - Notebooks
- Google Colab
- Kaggle
Sync ekko-pnc model card and cover
Browse files
README.md
CHANGED
|
@@ -128,6 +128,15 @@ device validation remains pending. PnC does not require a GPU.
|
|
| 128 |
|
| 129 |
## Training and export provenance
|
| 130 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 131 |
The shipped inference files are revision-pinned and checksum-verified. The
|
| 132 |
public [PnC recipe](https://github.com/Rye-A1/ekko-danish-stt/blob/main/training/recipes/pnc.toml)
|
| 133 |
preserves the final two-stage lineage and retained hyperparameters. The ONNX
|
|
|
|
| 128 |
|
| 129 |
## Training and export provenance
|
| 130 |
|
| 131 |
+
| Stage | Data used | Purpose |
|
| 132 |
+
|---|---|---|
|
| 133 |
+
| Stage 1 | Danish Wikipedia (`20231101.da`), 69,820,413 word labels | Learn broad Danish punctuation and casing from text |
|
| 134 |
+
| Stage 2 | NST-da, CoRal conversation/read-aloud, and FT Speech transcript text, 9,974,406 word labels | Adapt punctuation and casing to ASR-style text |
|
| 135 |
+
| Fused export | No new training data | Fold the final heads into one 10-class model and export safetensors plus fp32/int8 ONNX |
|
| 136 |
+
|
| 137 |
+
Nota, Common Voice, and FLEURS were not part of the retained PnC training
|
| 138 |
+
lineage.
|
| 139 |
+
|
| 140 |
The shipped inference files are revision-pinned and checksum-verified. The
|
| 141 |
public [PnC recipe](https://github.com/Rye-A1/ekko-danish-stt/blob/main/training/recipes/pnc.toml)
|
| 142 |
preserves the final two-stage lineage and retained hyperparameters. The ONNX
|