Upload folder using huggingface_hub
Browse files- README.md +82 -0
- config.json +13 -0
- model.pt +3 -0
- tokenizer.json +0 -0
- tokenizer_config.json +24 -0
README.md
ADDED
|
@@ -0,0 +1,82 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- ru
|
| 4 |
+
license: mit
|
| 5 |
+
pipeline_tag: text-classification
|
| 6 |
+
tags:
|
| 7 |
+
- toxicity
|
| 8 |
+
- multi-task
|
| 9 |
+
- rubert
|
| 10 |
+
- moderation
|
| 11 |
+
metrics:
|
| 12 |
+
- f1
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
+
# Multi-Task Toxicity Classification Model (ruBERT-tiny2)
|
| 16 |
+
|
| 17 |
+
Данная модель предназначена для автоматической модерации русскоязычных текстов и одновременно решает 3 задачи классификации:
|
| 18 |
+
1. **Profanity** — ненормативная лексика и мат.
|
| 19 |
+
2. **Threat** — угрозы насилия и физической расправы.
|
| 20 |
+
3. **Illegal** — обсуждение/призывы к противоправным и запрещенным действиям.
|
| 21 |
+
|
| 22 |
+
Основана на лёгком энкодере `cointegrated/rubert-tiny2` с тремя независимыми классификационными головами.
|
| 23 |
+
|
| 24 |
+
## 📊 Метрики на Валидации
|
| 25 |
+
|
| 26 |
+
| Категория | Precision | Recall | F1-Score | Оптимальный порог |
|
| 27 |
+
| :--- | :---: | :---: | :---: | :---: |
|
| 28 |
+
| **Profanity** | 0.8484 | 0.8534 | 0.8509 | 0.90 |
|
| 29 |
+
| **Threat** | 0.6532 | 0.8526 | 0.7397 | 0.85 |
|
| 30 |
+
| **Illegal** | 0.0000 | 0.0000 | 0.0000 | 0.50 |
|
| 31 |
+
|
| 32 |
+
**Итоговый Macro F1-Score:** 0.5302
|
| 33 |
+
|
| 34 |
+
## 🚀 Пример использования (Python / PyTorch)
|
| 35 |
+
|
| 36 |
+
```python
|
| 37 |
+
import torch
|
| 38 |
+
import torch.nn as nn
|
| 39 |
+
from huggingface_hub import hf_hub_download
|
| 40 |
+
from transformers import AutoModel, AutoTokenizer
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
class MultiTaskToxicityEncoder(nn.Module):
|
| 44 |
+
|
| 45 |
+
def __init__(self, model_name="cointegrated/rubert-tiny2"):
|
| 46 |
+
super().__init__()
|
| 47 |
+
self.encoder = AutoModel.from_pretrained(model_name)
|
| 48 |
+
hidden_size = self.encoder.config.hidden_size
|
| 49 |
+
self.profanity_head = nn.Linear(hidden_size, 1)
|
| 50 |
+
self.threat_head = nn.Linear(hidden_size, 1)
|
| 51 |
+
self.illegal_head = nn.Linear(hidden_size, 1)
|
| 52 |
+
|
| 53 |
+
def forward(self, input_ids, attention_mask):
|
| 54 |
+
outputs = self.encoder(
|
| 55 |
+
input_ids=input_ids, attention_mask=attention_mask
|
| 56 |
+
)
|
| 57 |
+
cls_embedding = outputs.last_hidden_state[:, 0, :]
|
| 58 |
+
return {
|
| 59 |
+
"profanity": self.profanity_head(cls_embedding).squeeze(-1),
|
| 60 |
+
"threat": self.threat_head(cls_embedding).squeeze(-1),
|
| 61 |
+
"illegal": self.illegal_head(cls_embedding).squeeze(-1),
|
| 62 |
+
}
|
| 63 |
+
|
| 64 |
+
|
| 65 |
+
# Загрузка токенизатора и весов
|
| 66 |
+
repo_id = "fdlvaaa/rubert-tiny2-multitask-toxicity"
|
| 67 |
+
tokenizer = AutoTokenizer.from_pretrained(repo_id)
|
| 68 |
+
model = MultiTaskToxicityEncoder()
|
| 69 |
+
|
| 70 |
+
weights_path = hf_hub_download(repo_id=repo_id, filename="model.pt")
|
| 71 |
+
model.load_state_dict(torch.load(weights_path, map_location="cpu"))
|
| 72 |
+
model.eval()
|
| 73 |
+
|
| 74 |
+
# Инференс
|
| 75 |
+
text = "Ты еще пожалеешь, я найду тебя!"
|
| 76 |
+
inputs = tokenizer(text, return_tensors="pt")
|
| 77 |
+
|
| 78 |
+
with torch.no_grad():
|
| 79 |
+
outputs = model(inputs["input_ids"], inputs["attention_mask"])
|
| 80 |
+
probs = {k: torch.sigmoid(v).item() for k, v in outputs.items()}
|
| 81 |
+
|
| 82 |
+
print("Вероятности:", probs)
|
config.json
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model_name": "cointegrated/rubert-tiny2",
|
| 3 |
+
"tasks": [
|
| 4 |
+
"profanity",
|
| 5 |
+
"threat",
|
| 6 |
+
"illegal"
|
| 7 |
+
],
|
| 8 |
+
"thresholds": {
|
| 9 |
+
"profanity": 0.9000000000000002,
|
| 10 |
+
"threat": 0.8500000000000002,
|
| 11 |
+
"illegal": 0.5
|
| 12 |
+
}
|
| 13 |
+
}
|
model.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:900a2fbd84d284648fa92a5d77ddf59e4f617be4e264a58561baa050c3f9ebd6
|
| 3 |
+
size 116801254
|
tokenizer.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
tokenizer_config.json
ADDED
|
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"backend": "tokenizers",
|
| 3 |
+
"cls_token": "[CLS]",
|
| 4 |
+
"do_basic_tokenize": true,
|
| 5 |
+
"do_lower_case": false,
|
| 6 |
+
"is_local": false,
|
| 7 |
+
"local_files_only": false,
|
| 8 |
+
"mask_token": "[MASK]",
|
| 9 |
+
"max_length": 512,
|
| 10 |
+
"model_max_length": 2048,
|
| 11 |
+
"never_split": null,
|
| 12 |
+
"pad_to_multiple_of": null,
|
| 13 |
+
"pad_token": "[PAD]",
|
| 14 |
+
"pad_token_type_id": 0,
|
| 15 |
+
"padding_side": "right",
|
| 16 |
+
"sep_token": "[SEP]",
|
| 17 |
+
"stride": 0,
|
| 18 |
+
"strip_accents": null,
|
| 19 |
+
"tokenize_chinese_chars": true,
|
| 20 |
+
"tokenizer_class": "BertTokenizer",
|
| 21 |
+
"truncation_side": "right",
|
| 22 |
+
"truncation_strategy": "longest_first",
|
| 23 |
+
"unk_token": "[UNK]"
|
| 24 |
+
}
|