detoxify

Runtime error

App Files Files Community

Utiric commited on Mar 13

Commit

ebef9a1

1 Parent(s): c7133b5

Changed usage

Browse files

Files changed (1) hide show

app.py +70 -12

app.py CHANGED Viewed

@@ -1,20 +1,21 @@
 from flask import Flask, request, jsonify, render_template_string
 import os
 import torch
 from detoxify import Detoxify
 from transformers import AutoModelForSequenceClassification, AutoTokenizer
 app = Flask(__name__)
-# Load models
 detoxify_model = Detoxify('multilingual')
 koala_model = AutoModelForSequenceClassification.from_pretrained("KoalaAI/Text-Moderation")
 koala_tokenizer = AutoTokenizer.from_pretrained("KoalaAI/Text-Moderation")
-# API key from environment variable
 API_KEY = os.getenv('API_KEY')
-# Modern HTML interface with Tailwind CSS
 HTML_TEMPLATE = '''
 <!DOCTYPE html>
 <html lang="en">
@@ -55,7 +56,7 @@ HTML_TEMPLATE = '''
           method: 'POST',
           headers: {
               'Content-Type': 'application/json',
-              'Authorization': 'Bearer YOUR_API_KEY'  // Replace YOUR_API_KEY with your actual API key
           },
           body: JSON.stringify({ model: model, texts: [text] })
         });
@@ -67,10 +68,11 @@ HTML_TEMPLATE = '''
           let html = '<h2 class="text-2xl font-bold mb-4">Results:</h2>';
           data.results.forEach(item => {
             html += `<div class="mb-4 p-4 bg-gray-200 dark:bg-gray-700 rounded">
-                       <p class="font-semibold">Input: ${item.input}</p>
                        <ul>`;
-            for (const [key, value] of Object.entries(item.predictions)) {
-              html += `<li>${key}: ${value.toFixed(5)}</li>`;
             }
             html += `  </ul>
                      </div>`;
@@ -86,13 +88,53 @@ HTML_TEMPLATE = '''
 </html>
 '''
 @app.route('/')
 def home():
     return render_template_string(HTML_TEMPLATE)
 @app.route('/v1/moderations', methods=['POST'])
 def moderations():
-    # Retrieve API key from Authorization header in Bearer format
     auth_header = request.headers.get('Authorization')
     if not auth_header or not auth_header.startswith("Bearer "):
         return jsonify({"error": "Unauthorized"}), 401
@@ -108,29 +150,45 @@ def moderations():
         return jsonify({"error": "Invalid input, expected a list of texts"}), 400
     results = []
     if model_choice == "koalaai/text-moderation":
         for text in texts:
             inputs = koala_tokenizer(text, return_tensors="pt")
             outputs = koala_model(**inputs)
             logits = outputs.logits
             probabilities = torch.softmax(logits, dim=-1).squeeze().tolist()
             if isinstance(probabilities, float):
                 probabilities = [probabilities]
             labels = [koala_model.config.id2label[idx] for idx in range(len(probabilities))]
             prediction = {label: prob for label, prob in zip(labels, probabilities)}
-            results.append({"input": text, "predictions": prediction})
         response_model = "koalaai/text-moderation"
     else:
         for text in texts:
             pred = detoxify_model.predict([text])
             prediction = {k: v[0] for k, v in pred.items()}
-            results.append({"input": text, "predictions": prediction})
         response_model = "unitaryai/detoxify-multilingual"
     response_data = {
-        "object": "moderation",
         "model": response_model,
-        "results": results
     }
     return jsonify(response_data)

 from flask import Flask, request, jsonify, render_template_string
 import os
+import uuid
 import torch
 from detoxify import Detoxify
 from transformers import AutoModelForSequenceClassification, AutoTokenizer
 app = Flask(__name__)
+# Modelleri yükle
 detoxify_model = Detoxify('multilingual')
 koala_model = AutoModelForSequenceClassification.from_pretrained("KoalaAI/Text-Moderation")
 koala_tokenizer = AutoTokenizer.from_pretrained("KoalaAI/Text-Moderation")
+# API key environment variable'dan
 API_KEY = os.getenv('API_KEY')
+# Modern, TailwindCSS destekli HTML arayüzü (dark/light)
 HTML_TEMPLATE = '''
 <!DOCTYPE html>
 <html lang="en">
           method: 'POST',
           headers: {
               'Content-Type': 'application/json',
+              'Authorization': 'Bearer YOUR_API_KEY'  // Değiştir!
           },
           body: JSON.stringify({ model: model, texts: [text] })
         });
           let html = '<h2 class="text-2xl font-bold mb-4">Results:</h2>';
           data.results.forEach(item => {
             html += `<div class="mb-4 p-4 bg-gray-200 dark:bg-gray-700 rounded">
+                       <p class="font-semibold">Flagged: ${item.flagged}</p>
+                       <p class="font-semibold">Categories:</p>
                        <ul>`;
+            for (const [key, value] of Object.entries(item.categories)) {
+              html += `<li>${key}: ${value} (score: ${item.category_scores[key].toFixed(5)})</li>`;
             }
             html += `  </ul>
                      </div>`;
 </html>
 '''
+def transform_predictions(model_choice, prediction_dict):
+    """
+    OpenAI tarzı response formatı için tahminleri dönüştür.
+    Dönüştürülen alanlar: flagged, categories, category_scores, category_applied_input_types
+    """
+    # OpenAI'nın moderasyon API'sinde kullanılan kategori anahtarları
+    category_keys = [
+        "sexual", "sexual/minors", "harassment", "harassment/threatening",
+        "hate", "hate/threatening", "illicit", "illicit/violent",
+        "self-harm", "self-harm/intent", "self-harm/instructions",
+        "violence", "violence/graphic"
+    ]
+    scores = {}
+    # Detoxify modeli için örnek eşleştirme:
+    if model_choice == "unitaryai/detoxify-multilingual":
+        scores["sexual"] = prediction_dict.get("sexual_explicit", 0.0)
+        scores["harassment"] = max(prediction_dict.get("identity_attack", 0.0), prediction_dict.get("insult", 0.0))
+        scores["hate"] = prediction_dict.get("toxicity", 0.0)
+        scores["violence"] = max(prediction_dict.get("severe_toxicity", 0.0), prediction_dict.get("threat", 0.0))
+        # Diğer kategoriler için varsayılan 0 değeri
+        for key in category_keys:
+            if key not in scores:
+                scores[key] = 0.0
+    else:
+        # Koala modelinden gelen tahminlerde, label isimleri doğrudan uyumlu olabilir;
+        # uyumlu değilse varsayılan 0 değeri ver.
+        for key in category_keys:
+            scores[key] = prediction_dict.get(key, 0.0)
+    # Kategori boolean değerleri; eşik değeri: 0.7
+    threshold = 0.7
+    bool_categories = {key: (scores[key] > threshold) for key in category_keys}
+    # category_applied_input_types: biz text ile çalıştığımız için, skor > 0 ise ["text"] değilse boş liste
+    cat_applied_input_types = {key: (["text"] if scores[key] > 0 else []) for key in category_keys}
+    # Flagged: herhangi bir kategori eşik değerinin üzerinde ise True
+    flagged = any(bool_categories.values())
+    return flagged, bool_categories, scores, cat_applied_input_types
 @app.route('/')
 def home():
     return render_template_string(HTML_TEMPLATE)
 @app.route('/v1/moderations', methods=['POST'])
 def moderations():
+    # Authorization header'dan API key kontrolü
     auth_header = request.headers.get('Authorization')
     if not auth_header or not auth_header.startswith("Bearer "):
         return jsonify({"error": "Unauthorized"}), 401
         return jsonify({"error": "Invalid input, expected a list of texts"}), 400
     results = []
+    # Her bir metin için tahmin ve transform işlemi
     if model_choice == "koalaai/text-moderation":
         for text in texts:
             inputs = koala_tokenizer(text, return_tensors="pt")
             outputs = koala_model(**inputs)
             logits = outputs.logits
             probabilities = torch.softmax(logits, dim=-1).squeeze().tolist()
+            # Eğer tek değer ise listeye çevir
             if isinstance(probabilities, float):
                 probabilities = [probabilities]
             labels = [koala_model.config.id2label[idx] for idx in range(len(probabilities))]
             prediction = {label: prob for label, prob in zip(labels, probabilities)}
+            flagged, bool_categories, scores, cat_applied_input_types = transform_predictions(model_choice, prediction)
+            results.append({
+                "flagged": flagged,
+                "categories": bool_categories,
+                "category_scores": scores,
+                "category_applied_input_types": cat_applied_input_types
+            })
         response_model = "koalaai/text-moderation"
     else:
         for text in texts:
             pred = detoxify_model.predict([text])
+            # Detoxify sonuçları liste formatında, tek değer alıyoruz
             prediction = {k: v[0] for k, v in pred.items()}
+            flagged, bool_categories, scores, cat_applied_input_types = transform_predictions(model_choice, prediction)
+            results.append({
+                "flagged": flagged,
+                "categories": bool_categories,
+                "category_scores": scores,
+                "category_applied_input_types": cat_applied_input_types
+            })
         response_model = "unitaryai/detoxify-multilingual"
     response_data = {
+        "id": "modr-" + uuid.uuid4().hex[:24],
         "model": response_model,
+        "results": results,
+        "object": "moderation"
     }
     return jsonify(response_data)