|
--- |
|
language: |
|
- ru |
|
tags: |
|
- text-classification |
|
license: apache-2.0 |
|
|
|
--- |
|
|
|
# RuBERTConv Toxic Classifier |
|
|
|
## Model description |
|
|
|
Based on [rubert-base-cased-conversational](https://huggingface.co/DeepPavlov/rubert-base-cased-conversational) model |
|
|
|
## Intended uses & limitations |
|
|
|
#### How to use |
|
|
|
Colab: [link](https://colab.research.google.com/drive/1veKO9hke7myxKigZtZho_F-UM2fD9kp8) |
|
|
|
```python |
|
from transformers import pipeline |
|
|
|
model_name = "IlyaGusev/rubertconv_toxic_clf" |
|
pipe = pipeline("text-classification", model=model_name, tokenizer=model_name, framework="pt") |
|
|
|
text = "Ты придурок из интернета" |
|
pipe([text]) |
|
``` |
|
|
|
## Training data |
|
|
|
Datasets: |
|
- [2ch]( https://www.kaggle.com/blackmoon/russian-language-toxic-comments) |
|
- [Odnoklassniki](https://www.kaggle.com/alexandersemiletov/toxic-russian-comments) |
|
- [Toloka Persona Chat Rus](https://toloka.ai/ru/datasets) |
|
- [Koziev's Conversations](https://github.com/Koziev/NLP_Datasets/blob/master/Conversations/Data) with [toxic words vocabulary](https://www.dropbox.com/s/ou6lx03b10yhrfl/bad_vocab.txt.tar.gz) |
|
|
|
Augmentations: |
|
- ё -> е |
|
- Remove or add "?" or "!" |
|
- Fix CAPS |
|
- Concatenate toxic and non-toxic texts |
|
- Concatenate two non-toxic texts |
|
- Add toxic words from vocabulary |
|
- Add typos |
|
- Mask toxic words with "*", "@", "$" |
|
|
|
|
|
## Training procedure |
|
|
|
TBA |