Update app.py

app.py

@@ -2,23 +2,25 @@ import gradio as gr
 import os
 from huggingface_hub import InferenceClient
 
-# Load token and model - Try different models if one fails
+# HF Token aus Umgebungsvariable laden
 HF_TOKEN = os.getenv("tomoniaccess")
 
-# Model options (uncomment one that works):
-model_name = "LeoLM/leo-hessianai-13b-chat"        # Primary choice
+# Modellname definieren
+model_name = "LeoLM/leo-hessianai-13b-chat"
+
+# Client initialisieren
 client = InferenceClient(
     model=model_name,
     token=HF_TOKEN
 )
 
-# Initialize conversation history as global variable
+# Globale Konversationshistorie
 conversation_history = []
 
 def query_leolm(messages, max_tokens=200, temperature=1.0, top_p=0.9):
-    """Query LeoLM-13B model via Hugging Face InferenceClient"""
+    """Versuche Chat-Completion, falle auf Text-Generation zurück."""
     try:
-        # Try chat completion first
+        # Versuch über chat_completion
         response = client.chat_completion(
             messages=messages,
             max_tokens=max_tokens,
@@ -30,39 +32,38 @@ def query_leolm(messages, max_tokens=200, temperature=1.0, top_p=0.9):
     except Exception as chat_error:
         print(f"Chat completion failed: {chat_error}")
         try:
-            # Fallback to text generation if chat completion fails
-            # Convert messages to single prompt
+            # Fallback: Prompt manuell zusammensetzen
             prompt = ""
             for msg in messages:
                 if msg["role"] == "system":
-                    prompt += f"System: {msg['content']}\n\n"
+                    prompt += f"<|system|>\n{msg['content'].strip()}\n"
                 elif msg["role"] == "user":
-                    prompt += f"User: {msg['content']}\n"
+                    prompt += f"<|user|>\n{msg['content'].strip()}\n"
                 elif msg["role"] == "assistant":
-                    prompt += f"Assistant: {msg['content']}\n"
-            
-            prompt += "Assistant: "
-            
+                    prompt += f"<|assistant|>\n{msg['content'].strip()}\n"
+            prompt += "<|assistant|>\n"
+
             response = client.text_generation(
                 prompt=prompt,
                 max_new_tokens=max_tokens,
                 temperature=temperature,
                 top_p=top_p,
                 do_sample=True,
+                repetition_penalty=1.1,
+                stop_sequences=["<|user|>", "<|system|>"],
                 return_full_text=False
             )
-            return response
+            return response.strip()
         except Exception as text_error:
             print(f"Text generation also failed: {text_error}")
             raise Exception(f"Both methods failed - Chat: {str(chat_error)}, Text: {str(text_error)}")
 
 def enhanced_chat_response(user_input, max_tokens, temperature, top_p):
-    global conversation_history  # Declare global variable
-    
+    global conversation_history
+
     if not user_input.strip():
         return "", "*Bitte gib eine Nachricht ein.*", ""
 
-    # Optimized system prompt for LeoLM (German-focused model)
     system_prompt = {
         "role": "system",
         "content": (
@@ -73,7 +74,7 @@ def enhanced_chat_response(user_input, max_tokens, temperature, top_p):
         )
     }
 
-    # Keep last 6 messages for context (LeoLM works well with moderate context)
+    # Kürze History falls nötig
     messages = [system_prompt] + conversation_history[-6:]
     current_message = {"role": "user", "content": user_input}
     messages.append(current_message)
@@ -100,129 +101,64 @@ def reset_conversation():
     return "Neues Gespräch gestartet.", ""
 
 def test_api_connection():
-    """Test API connection with multiple fallback methods"""
     try:
-        # Test 1: Simple chat completion
         test_messages = [
-            {"role": "system", "content": "Du bist ein hilfsbereit Assistent und antwortest auf Deutsch."},
+            {"role": "system", "content": "Du bist ein Assistent."},
             {"role": "user", "content": "Hallo"}
         ]
-        
         response = query_leolm(test_messages, max_tokens=20)
         return f"✅ API Verbindung erfolgreich: {response[:50]}..."
     except Exception as e:
-        # Test 2: Try direct text generation
         try:
             simple_response = client.text_generation(
                 prompt="Hallo, wie geht es dir?",
                 max_new_tokens=10,
-                do_sample=False,
                 return_full_text=False
             )
             return f"✅ API Verbindung (Text Generation): {simple_response[:50]}..."
         except Exception as e2:
-            # Test 3: Check if model exists
-            try:
-                # Try to get model info
-                model_info = f"Model: {model_name}"
-                return f"❌ API Errors - Chat: {str(e)[:100]}... | Text: {str(e2)[:100]}... | {model_info}"
-            except Exception as e3:
-                return f"❌ Vollständiger API Fehler: {str(e)[:200]}..."
-
-# UI
+            return f"❌ Fehler: {str(e)[:100]} | {str(e2)[:100]}"
+
+# Gradio UI
 with gr.Blocks(title="LeoLM Depression Training Simulator") as demo:
     gr.Markdown("## 🧠 Depression Training Simulator (LeoLM-13B)")
-    gr.Markdown("**Übe realistische Gespräche mit einem 16-jährigen Teenager mit Depressionen.**")
-    gr.Markdown("*Powered by LeoLM/leo-hessianai-13b-chat - Deutsches Sprachmodell*")
+    gr.Markdown("**Simuliere Gespräche mit einem 16-jährigen Teenager mit Depressionen.**")
+    gr.Markdown("*Sprachmodell: `LeoLM/leo-hessianai-13b-chat`*")
 
     with gr.Row():
         with gr.Column(scale=1):
-            gr.Markdown("### ⚙️ Einstellungen")
             max_tokens = gr.Slider(50, 300, value=150, step=10, label="Max. Antwortlänge")
             temperature = gr.Slider(0.1, 1.5, value=0.8, step=0.1, label="Kreativität (Temperature)")
             top_p = gr.Slider(0.1, 1.0, value=0.9, step=0.05, label="Top-p (Fokus)")
 
-            gr.Markdown("### 🔧 API Status")
             api_status = gr.Textbox(label="Status", value="")
             api_test_btn = gr.Button("API testen")
-
-            gr.Markdown("### 🔄 Aktionen")
             reset_btn = gr.Button("Neues Gespräch")
-            
-            gr.Markdown("### 📋 Setup & Troubleshooting")
-            gr.Markdown("""
-            **Benötigt:**
-            - `tomoniaccess` Umgebungsvariable mit HF Token
-            - `pip install huggingface_hub gradio`
-            
-            **LeoLM Info:**
-            - Deutsche Sprachoptimierung
-            - 13B Parameter
-            - Modell: `LeoLM/leo-hessianai-13b-chat`
-            
-            **Bei API Fehlern:**
-            1. Token prüfen (muss Pro/Enterprise sein)
-            2. Modell verfügbar? → [HF Model Card](https://huggingface.co/LeoLM/leo-hessianai-13b-chat)
-            3. Alternative: `LeoLM/leo-hessianai-7b-chat`
-            4. Fallback: `microsoft/DialoGPT-medium`
-            """)
 
         with gr.Column(scale=2):
-            gr.Markdown("### 💬 Gespräch")
-            user_input = gr.Textbox(
-                label="Deine Nachricht", 
-                placeholder="Hallo, wie geht es dir heute?",
-                lines=2
-            )
+            user_input = gr.Textbox(label="Deine Nachricht", placeholder="Wie fühlst du dich heute?", lines=2)
             send_btn = gr.Button("📨 Senden")
+            bot_response = gr.Textbox(label="Antwort", value="", lines=3)
+            chat_history = gr.Textbox(label="Gesprächsverlauf", value="", lines=15)
 
-            bot_response = gr.Textbox(
-                label="Antwort", 
-                value="",
-                lines=3
-            )
+    send_btn.click(fn=enhanced_chat_response,
+                   inputs=[user_input, max_tokens, temperature, top_p],
+                   outputs=[user_input, bot_response, chat_history])
 
-            chat_history = gr.Textbox(
-                label="Gesprächsverlauf",
-                value="",
-                lines=15
-            )
+    user_input.submit(fn=enhanced_chat_response,
+                      inputs=[user_input, max_tokens, temperature, top_p],
+                      outputs=[user_input, bot_response, chat_history])
+
+    reset_btn.click(fn=reset_conversation,
+                    outputs=[bot_response, chat_history])
 
-    # Event Bindings
-    send_btn.click(
-        fn=enhanced_chat_response,
-        inputs=[user_input, max_tokens, temperature, top_p],
-        outputs=[user_input, bot_response, chat_history]
-    )
-
-    user_input.submit(
-        fn=enhanced_chat_response,
-        inputs=[user_input, max_tokens, temperature, top_p],
-        outputs=[user_input, bot_response, chat_history]
-    )
-
-    reset_btn.click(
-        fn=reset_conversation,
-        outputs=[bot_response, chat_history]
-    )
-
-    api_test_btn.click(
-        fn=test_api_connection,
-        outputs=[api_status]
-    )
+    api_test_btn.click(fn=test_api_connection,
+                       outputs=[api_status])
 
 if __name__ == "__main__":
-    print("🚀 LeoLM Depression Training Simulator")
-    print(f"📊 Model: {model_name}")
-    
+    print("🚀 Starte LeoLM Depression Simulator")
     if not HF_TOKEN:
-        print("❌ FEHLER: tomoniaccess Umgebungsvariable ist nicht gesetzt!")
-        print("   Bitte setze deinen Hugging Face Token als 'tomoniaccess' Umgebungsvariable.")
+        print("❌ Umgebungsvariable 'tomoniaccess' nicht gesetzt.")
     else:
-        print("✅ Hugging Face API Token gefunden")
-    
-    print("\n📦 Benötigte Pakete:")
-    print("pip install huggingface_hub gradio")
-    print("\n🇩🇪 LeoLM: Deutsches Sprachmodell für bessere Konversationen")
-    
-    demo.launch(share=False)
+        print("✅ Token erkannt")
+    demo.launch()