Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
|
@@ -3,26 +3,17 @@ import re
|
|
| 3 |
import os
|
| 4 |
import torch
|
| 5 |
from transformers import BertTokenizer, AutoModelForSequenceClassification
|
| 6 |
-
|
| 7 |
-
# استدعاء معالج التنظيف الرسمي لـ AraBERT الخاص بكِ
|
| 8 |
from arabert.preprocess import ArabertPreprocessor
|
| 9 |
|
| 10 |
-
# معرفات المستودع والمسار الفرعي لأسماء
|
| 11 |
MODEL_REPO = "kkAsmaa/ChildShield"
|
| 12 |
MODEL_NAME = "aubmindlab/bert-base-arabertv02-twitter"
|
| 13 |
SUB_FOLDER = "ChildShield"
|
| 14 |
-
|
| 15 |
-
# سحب المفتاح السري تلقائياً وبأعلى درجات الأمان السيبراني
|
| 16 |
HF_TOKEN = os.getenv("HF_TOKEN")
|
| 17 |
-
|
| 18 |
print("🔄 Loading model weights from the secured ChildShield subfolder...")
|
| 19 |
|
| 20 |
-
# استدعاء المترجم المستقر وتوجيه الموديل بدقة للمجلد الفرعي
|
| 21 |
tokenizer = BertTokenizer.from_pretrained(MODEL_NAME)
|
| 22 |
model = AutoModelForSequenceClassification.from_pretrained(MODEL_REPO, token=HF_TOKEN, subfolder=SUB_FOLDER)
|
| 23 |
model.eval()
|
| 24 |
-
|
| 25 |
-
# تفعيل المعالج ليتطابق مع داتا كولاب
|
| 26 |
arabic_prep = ArabertPreprocessor(model_name=MODEL_NAME)
|
| 27 |
|
| 28 |
def clean_obfuscation(text):
|
|
@@ -35,7 +26,6 @@ def clean_obfuscation(text):
|
|
| 35 |
text = re.sub(r'[^\w\s\.]', ' ', text)
|
| 36 |
text = re.sub(r'\s+', ' ', text)
|
| 37 |
return text.strip()
|
| 38 |
-
|
| 39 |
def full_preprocess(text):
|
| 40 |
text_no_trickery = clean_obfuscation(text)
|
| 41 |
final_text = arabic_prep.preprocess(text_no_trickery)
|
|
@@ -43,16 +33,13 @@ def full_preprocess(text):
|
|
| 43 |
|
| 44 |
def predict_safety_api(text):
|
| 45 |
"""
|
| 46 |
-
|
| 47 |
-
and standardized native padding execution.
|
| 48 |
"""
|
| 49 |
cleaned_text = full_preprocess(text)
|
| 50 |
|
| 51 |
-
# 1. تقطيع النص الأولي لأرقام مجهولة الأبعاد بدون حشو مسبق
|
| 52 |
full_encodings = tokenizer(cleaned_text, add_special_tokens=False, return_attention_mask=False)
|
| 53 |
input_ids = full_encodings['input_ids']
|
| 54 |
|
| 55 |
-
# تثبيت أبعاد النوافذ الذكية الخاصة بكِ (60/20)
|
| 56 |
window_size = 60
|
| 57 |
overlap = 20
|
| 58 |
windows = []
|
|
@@ -70,7 +57,7 @@ def predict_safety_api(text):
|
|
| 70 |
highest_unsafe_prob = 0.0
|
| 71 |
|
| 72 |
for win_ids in windows:
|
| 73 |
-
|
| 74 |
window_text = tokenizer.decode(win_ids, skip_special_tokens=True)
|
| 75 |
|
| 76 |
inputs = tokenizer(
|
|
@@ -86,7 +73,6 @@ def predict_safety_api(text):
|
|
| 86 |
|
| 87 |
probs = torch.softmax(outputs.logits, dim=-1).flatten().tolist()
|
| 88 |
|
| 89 |
-
# 🎯 السطر الفائز المصلح: قراءة الخانة رقم 1 المخصصة لنسبة الخطر بدقة
|
| 90 |
unsafe_p = float(probs[1])
|
| 91 |
|
| 92 |
if unsafe_p > 0.50:
|
|
@@ -99,7 +85,6 @@ def predict_safety_api(text):
|
|
| 99 |
safe_p = 1.0 - highest_unsafe_prob
|
| 100 |
return {"verdict": "SAFE", "block": False, "confidence": f"{safe_p * 100:.2f}%"}
|
| 101 |
|
| 102 |
-
# بناء واجهة Gradio الاحترافية للمشروع
|
| 103 |
interface = gr.Interface(
|
| 104 |
fn=predict_safety_api,
|
| 105 |
inputs=gr.Textbox(lines=3, placeholder="Enter text to analyze..."),
|
|
|
|
| 3 |
import os
|
| 4 |
import torch
|
| 5 |
from transformers import BertTokenizer, AutoModelForSequenceClassification
|
|
|
|
|
|
|
| 6 |
from arabert.preprocess import ArabertPreprocessor
|
| 7 |
|
|
|
|
| 8 |
MODEL_REPO = "kkAsmaa/ChildShield"
|
| 9 |
MODEL_NAME = "aubmindlab/bert-base-arabertv02-twitter"
|
| 10 |
SUB_FOLDER = "ChildShield"
|
|
|
|
|
|
|
| 11 |
HF_TOKEN = os.getenv("HF_TOKEN")
|
|
|
|
| 12 |
print("🔄 Loading model weights from the secured ChildShield subfolder...")
|
| 13 |
|
|
|
|
| 14 |
tokenizer = BertTokenizer.from_pretrained(MODEL_NAME)
|
| 15 |
model = AutoModelForSequenceClassification.from_pretrained(MODEL_REPO, token=HF_TOKEN, subfolder=SUB_FOLDER)
|
| 16 |
model.eval()
|
|
|
|
|
|
|
| 17 |
arabic_prep = ArabertPreprocessor(model_name=MODEL_NAME)
|
| 18 |
|
| 19 |
def clean_obfuscation(text):
|
|
|
|
| 26 |
text = re.sub(r'[^\w\s\.]', ' ', text)
|
| 27 |
text = re.sub(r'\s+', ' ', text)
|
| 28 |
return text.strip()
|
|
|
|
| 29 |
def full_preprocess(text):
|
| 30 |
text_no_trickery = clean_obfuscation(text)
|
| 31 |
final_text = arabic_prep.preprocess(text_no_trickery)
|
|
|
|
| 33 |
|
| 34 |
def predict_safety_api(text):
|
| 35 |
"""
|
| 36 |
+
Arabic text classification gateway utilizing a custom sliding window configuration with 20 token overlap.
|
|
|
|
| 37 |
"""
|
| 38 |
cleaned_text = full_preprocess(text)
|
| 39 |
|
|
|
|
| 40 |
full_encodings = tokenizer(cleaned_text, add_special_tokens=False, return_attention_mask=False)
|
| 41 |
input_ids = full_encodings['input_ids']
|
| 42 |
|
|
|
|
| 43 |
window_size = 60
|
| 44 |
overlap = 20
|
| 45 |
windows = []
|
|
|
|
| 57 |
highest_unsafe_prob = 0.0
|
| 58 |
|
| 59 |
for win_ids in windows:
|
| 60 |
+
|
| 61 |
window_text = tokenizer.decode(win_ids, skip_special_tokens=True)
|
| 62 |
|
| 63 |
inputs = tokenizer(
|
|
|
|
| 73 |
|
| 74 |
probs = torch.softmax(outputs.logits, dim=-1).flatten().tolist()
|
| 75 |
|
|
|
|
| 76 |
unsafe_p = float(probs[1])
|
| 77 |
|
| 78 |
if unsafe_p > 0.50:
|
|
|
|
| 85 |
safe_p = 1.0 - highest_unsafe_prob
|
| 86 |
return {"verdict": "SAFE", "block": False, "confidence": f"{safe_p * 100:.2f}%"}
|
| 87 |
|
|
|
|
| 88 |
interface = gr.Interface(
|
| 89 |
fn=predict_safety_api,
|
| 90 |
inputs=gr.Textbox(lines=3, placeholder="Enter text to analyze..."),
|