|
|
--- |
|
|
language: |
|
|
- he |
|
|
license: cc-by-sa-4.0 |
|
|
tags: |
|
|
- text-classification |
|
|
- profanity-detection |
|
|
- hebrew |
|
|
- bert |
|
|
- alephbert |
|
|
library_name: transformers |
|
|
base_model: onlplab/alephbert-base |
|
|
datasets: |
|
|
- custom |
|
|
metrics: |
|
|
- accuracy |
|
|
- precision |
|
|
- recall |
|
|
- f1 |
|
|
--- |
|
|
|
|
|
# OpenCensor-Hebrew |
|
|
|
|
|
This is a fine tuned **AlephBERT** model that finds bad words ( profanity ) in Hebrew text. |
|
|
|
|
|
You give the model a Hebrew sentence. |
|
|
It returns: |
|
|
- a score between **0 and 1** |
|
|
- a yes/no flag (based on a cutoff you choose) |
|
|
|
|
|
Meaning of the score: |
|
|
- **0 = clean**, **1 = has profanity** |
|
|
- Recommended cutoff from tests: **0.29** ( you can change it ) |
|
|
|
|
|
 |
|
|
 |
|
|
 |
|
|
|
|
|
## How to use |
|
|
|
|
|
```python |
|
|
import torch |
|
|
from transformers import AutoTokenizer, AutoModelForSequenceClassification |
|
|
|
|
|
KModel = "LikoKIko/OpenCensor-Hebrew" |
|
|
KCutoff = 0.29 # best threshold from training |
|
|
KMaxLen = 256 # number of tokens (not characters) |
|
|
|
|
|
tokenizer = AutoTokenizer.from_pretrained(KModel) |
|
|
model = AutoModelForSequenceClassification.from_pretrained(KModel, num_labels=1).eval() |
|
|
|
|
|
text = "some hebrew text here" |
|
|
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=KMaxLen) |
|
|
|
|
|
with torch.inference_mode(): |
|
|
score = torch.sigmoid(model(**inputs).logits).item() |
|
|
KHasProfanity = int(score >= KCutoff) |
|
|
|
|
|
print({"score": round(score, 4), "KHasProfanity": KHasProfanity}) |
|
|
``` |
|
|
|
|
|
Note: If the text is very long, it is cut at `KMaxLen` tokens. |
|
|
|
|
|
## About this model |
|
|
|
|
|
- Base: `onlplab/alephbert-base` |
|
|
- Task: binary classification (clean / profanity) |
|
|
- Language: Hebrew |
|
|
- Max length: 256 tokens |
|
|
- Training (this run): |
|
|
- Batch size: 32 |
|
|
- Epochs: 10 |
|
|
- Learning rate: 0.000015 |
|
|
- Loss: binary cross-entropy with logits (`BCEWithLogitsLoss`). We use `pos_weight` so the model pays more attention to the rare class. This helps when the dataset is imbalanced. |
|
|
- Scheduler: linear warmup (10%) |
|
|
|
|
|
### Results (this run) |
|
|
|
|
|
- Test Accuracy: 0.9740 |
|
|
- Test Precision: 0.9726 |
|
|
- Test Recall: 0.9708 |
|
|
- Test F1: 0.9717 |
|
|
- Best threshold: 0.29 |
|
|
- ----------------------------- |
|
|
- *Updates: Set threshold to 0.8 |
|
|
|
|
|
## Reproduce (training code) |
|
|
|
|
|
This model was trained with a script that: |
|
|
- Loads `onlplab/alephbert-base` with `num_labels=1` |
|
|
- Tokenizes with `max_length=256` and pads to the max length |
|
|
- Trains with AdamW, linear warmup, and mixed precision |
|
|
- Tries cutoffs from `0.05` to `0.95` on the validation set and picks the best F1 |
|
|
- Saves the best checkpoint by validation F1, then reports test metrics |
|
|
|
|
|
## License |
|
|
|
|
|
CC-BY-SA-4.0 |
|
|
|
|
|
## How to cite |
|
|
|
|
|
```bibtex |
|
|
@misc{opencensor-hebrew, |
|
|
title = {OpenCensor-Hebrew: Hebrew Profanity Detection Model}, |
|
|
author = {LikoKIko}, |
|
|
year = {2025}, |
|
|
url = {https://huggingface.co/LikoKIko/OpenCensor-Hebrew} |
|
|
} |
|
|
``` |
|
|
|
