hlevring/bert-punct-restoration-da-onnx

Danish Punctuation & Capitalization Restoration (ONNX)

ONNX conversion of Alvenir/bert-punct-restoration-da for use with transformers.js and other ONNX runtimes.

Model Details

• Architecture: BertForTokenClassification (based on Maltehb/danish-bert-botxo)
• Parameters: ~110M
• Task: Token classification with 15 labels encoding punctuation + capitalization
• Language: Danish (da)

Labels

Each label has format <punctuation><case>:

Label	Punctuation	Capitalization
OO	None	lowercase
OU	None	Uppercase
.O	Period	lowercase
.U	Period	Uppercase
,O	Comma	lowercase
,U	Comma	Uppercase
?O	Question mark	lowercase
?U	Question mark	Uppercase
!O	Exclamation	lowercase
!U	Exclamation	Uppercase
:O	Colon	lowercase
:U	Colon	Uppercase
;O	Semicolon	lowercase
'O	Apostrophe	lowercase
-O	Hyphen	lowercase

Files

File	Precision	Size
onnx/model.onnx	FP32	~440 MB
onnx/model_fp16.onnx	FP16	~220 MB
onnx/model_quantized.onnx	INT8	~111 MB

Usage with transformers.js

import { AutoTokenizer, BertForTokenClassification } from "@huggingface/transformers";

const tokenizer = await AutoTokenizer.from_pretrained("hlevring/bert-punct-restoration-da-onnx");
const model = await BertForTokenClassification.from_pretrained("hlevring/bert-punct-restoration-da-onnx", {
  dtype: "q8",
});

const encoded = tokenizer("hej og velkommen til linket horoskop", { return_tensors: "pt" });
const output = await model(encoded);
// output.logits: [batch, seq_len, 15] - argmax to get label IDs

Available dtype options

• "fp32" - Full precision (440 MB)
• "fp16" - Half precision (220 MB)
• "q8" - INT8 quantized (111 MB)

Original Model

This is an ONNX export of Alvenir/bert-punct-restoration-da, part of the punctfix library.