Update README.md
Browse files
README.md
CHANGED
|
@@ -16,4 +16,49 @@ library_name: transformers
|
|
| 16 |
tags:
|
| 17 |
- moderation
|
| 18 |
- abuse_detection
|
| 19 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 16 |
tags:
|
| 17 |
- moderation
|
| 18 |
- abuse_detection
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
To use this model follow this.
|
| 22 |
+
|
| 23 |
+
# Load model directly
|
| 24 |
+
from transformers import AutoTokenizer, AutoModelForSequenceClassification
|
| 25 |
+
|
| 26 |
+
tokenizer = AutoTokenizer.from_pretrained("iamrazi/text-moderation")
|
| 27 |
+
model = AutoModelForSequenceClassification.from_pretrained("iamrazi/text-moderation")
|
| 28 |
+
|
| 29 |
+
model.eval() # Set model to evaluation mode
|
| 30 |
+
|
| 31 |
+
def predict_abuse(text: str, threshold: float = 0.5):
|
| 32 |
+
"""
|
| 33 |
+
Predict if a text is abusive or not.
|
| 34 |
+
|
| 35 |
+
Args:
|
| 36 |
+
text (str): Input text.
|
| 37 |
+
threshold (float): Probability threshold for classification.
|
| 38 |
+
|
| 39 |
+
Returns:
|
| 40 |
+
label (int): 0 for non-abusive, 1 for abusive
|
| 41 |
+
proba (float): Probability of being abusive
|
| 42 |
+
"""
|
| 43 |
+
# Tokenize
|
| 44 |
+
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=128)
|
| 45 |
+
|
| 46 |
+
# Forward pass
|
| 47 |
+
with torch.no_grad():
|
| 48 |
+
outputs = model(**inputs)
|
| 49 |
+
logits = outputs.logits
|
| 50 |
+
probas = torch.sigmoid(logits) # if your model output layer is logits
|
| 51 |
+
|
| 52 |
+
# For binary classification, take the probability of class 1
|
| 53 |
+
prob = probas[0][1].item() if probas.shape[1] > 1 else probas[0][0].item()
|
| 54 |
+
|
| 55 |
+
# Determine label
|
| 56 |
+
label = 1 if prob >= threshold else 0
|
| 57 |
+
|
| 58 |
+
return label, prob
|
| 59 |
+
|
| 60 |
+
|
| 61 |
+
text = "तुम बहुत गंदे हो 😡"
|
| 62 |
+
label, proba = predict_abuse(text)
|
| 63 |
+
|
| 64 |
+
Output: Label: 0, Probability: 0.08
|