README.md

---
license: apache-2.0
tags:
- generated_from_trainer
model-index:
- name: mt5_correct_puntuation_v3
  results: []
---

<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->

# mt5_correct_puntuation

本模型使用中文維基百科語料微調 [google/mt5-base](https://huggingface.co/google/mt5-base)預訓練模型之中文標點符號訂正器。目前之準確率為 0.794。

This is a [google/mt5-base](https://huggingface.co/google/mt5-base) model trained on Mandarin Wikipedia corpus and finetuned for Mandarin punctuation correction.  Currently the accuracy is 0.794.

## Datasets
模型使用中文維基百科公開資料微調。將取得的文本以「。」或「，」切分為不超過100字的句子。因為逗號和句號數量壓倒性地多，為盡量平衡資料集，僅保留包含冒號、分號、驚嘆號、問號的句子，作為正確句。將正確句之「，。：；、！？」隨機以「，。：；、！？」，製作為不正確句。訓練用句子共有291,112句。

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 5e-05
- train_batch_size: 4
- eval_batch_size: 4
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 1

### Framework versions

- Transformers 4.20.1
- Pytorch 1.12.0+cu113
- Datasets 2.3.2
- Tokenizers 0.12.1