Update app.py

app.py

@@ -10,7 +10,11 @@ from huggingface_hub.utils import EntryNotFoundError
 HF_TOKEN        = os.environ.get("HF_TOKEN", "")
 DATASET_REPO    = os.environ.get("DATASET_REPO", "")
 DATASET_FILE    = "data.jsonl"
-MODEL_ID        = "meta-llama/Llama-4-Maverick-17B-128E-Instruct"
+
+# Modelle aufteilen!
+MODEL_DEFAULT   = "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8"
+MODEL_TUNING    = "meta-llama/Llama-3.3-70B-Instruct"
+
 LEADERBOARD_PATH = os.path.join(os.path.dirname(os.path.abspath(__file__)), "leaderboard.json")
 
 # ── System Prompts ─────────────────────────────────────────────────────────────
@@ -315,10 +319,11 @@ _init_leaderboard()
 
 # ── LLM-Calls ─────────────────────────────────────────────────────────────────
 
-def _call_llm(system, user, max_tokens=1024):
+# Neu: Nimmt modell ID als Parameter an (Default ist Llama 4)
+def _call_llm(system, user, max_tokens=1024, model_id=MODEL_DEFAULT):
     client = InferenceClient(provider="novita", api_key=HF_TOKEN)
     resp = client.chat.completions.create(
-        model=MODEL_ID,
+        model=model_id,
         messages=[{"role": "system", "content": system},
                   {"role": "user",   "content": user}],
         max_tokens=max_tokens,
@@ -333,6 +338,7 @@ def translate(text, direction):
         return "Kein HF_TOKEN gefunden."
     prompt = PROMPT_TO_LINKEDIN if direction == "to_linkedin" else PROMPT_FROM_LINKEDIN
     try:
+        # Nutzt standardmäßig Llama 4
         return _call_llm(prompt, text)
     except Exception as e:
         return f"Fehler: {e}"
@@ -366,10 +372,22 @@ def get_bingo(text):
         try:
             raw     = _call_llm(PROMPT_BINGO, text, max_tokens=600)
             data    = _extract_json(raw)
-            metrics = data.get("metrics", [])
+            
+            # --- TÜRSTEHER FÜR DAS LLM-JSON ---
+            erlaubte_labels = ["Buzzword-Dichte", "Länge vs. Inhalt", "Selbstbeweihräuche", "Hashtag-Overload", "Sinnlosigkeits-Index"]
+            
+            # 1. Nur die 5 exakten Metriken erlauben
+            metrics = [m for m in data.get("metrics", []) if m.get("label") in erlaubte_labels]
+            data["metrics"] = metrics # Für den Renderer aktualisieren
+            
+            # 2. Score berechnen (jetzt garantiert max 50)
             total   = sum(int(m.get("score", 0)) for m in metrics)
             max_s   = len(metrics) * 10
-            verdict = data.get("verdict", "")
+            
+            # 3. Fallback für das Urteil
+            verdict = data.get("verdict", "Die KI war sprachlos: Kein Urteil generiert.")
+            data["verdict"] = verdict
+            # ----------------------------------
 
             _add_to_lb(text, total, max_s, verdict)
 
@@ -407,11 +425,56 @@ TUNING:
 - Zielgruppe: {zielgruppe}
 - Call to Action: {cta}"""
     try:
-        return _call_llm(prompt, user_msg, max_tokens=800)
+        # Hier nutzen wir das mächtige 70B Modell!
+        return _call_llm(prompt, user_msg, max_tokens=800, model_id=MODEL_TUNING)
     except Exception as e:
         return f"Fehler: {e}"
 
 
+# --- LADE ANIMATIONEN ---
+def generate_tuned_post_with_loader(original_text, ton, substanz, laenge, zielgruppe, cta):
+    if not original_text.strip():
+        yield "Bitte zuerst einen Post eingeben."
+        return
+        
+    yield "⏳ KI (Llama 3.3 70B) optimiert den Post nach deinen Vorgaben... Bitte warten."
+    result = generate_tuned_post(original_text, ton, substanz, laenge, zielgruppe, cta)
+    yield result
+
+
+def run_translate(text, direction):
+    if not text.strip():
+        yield "", gr.update(), gr.update(), gr.update()
+        return
+
+    if direction == "to_linkedin":
+        yield ("",
+               gr.update(value="⏳ **Generiere epische LinkedIn-Prosa...** Bitte warten.", visible=True),
+               gr.update(visible=False),
+               gr.update())
+               
+        result = translate(text, direction)
+        
+        yield (result,
+               gr.update(value=result, visible=True),
+               gr.update(visible=False),
+               gr.update())
+    else:
+        loader_html = '<div class="loading-box"><div class="spinner"></div><span>Analysiere Corporate Nonsense...</span></div>'
+        yield ("",
+               gr.update(visible=False),
+               gr.update(value=loader_html, visible=True),
+               gr.update())
+               
+        result = translate(text, direction)
+        bingo_html, lb_html = get_bingo(text)
+        
+        yield (result,
+               gr.update(visible=False),
+               gr.update(value=bingo_html, visible=True),
+               gr.update(value=lb_html))
+
+
 def _render_bingo(data):
     metrics = data.get("metrics", [])
     verdict = data.get("verdict", "")
@@ -490,23 +553,6 @@ def swap_direction(current_dir, inp, out):
     return new_dir, out, inp, *_labels(new_dir)
 
 
-# ── Haupt-Handler ──────────────────────────────────────────────────────────────
-
-def run_translate(text, direction):
-    result = translate(text, direction)
-    if direction == "to_linkedin":
-        return (result,
-                gr.update(value=result, visible=True),
-                gr.update(value="",     visible=False),
-                gr.update(value=_render_leaderboard()))
-    else:
-        bingo_html, lb_html = get_bingo(text)
-        return (result,
-                gr.update(value="",         visible=False),
-                gr.update(value=bingo_html, visible=True),
-                gr.update(value=lb_html))
-
-
 # ── CSS ───────────────────────────────────────────────────────────────────────
 
 CSS = """
@@ -578,6 +624,19 @@ button.secondary:hover { background:var(--li-blue-light) !important; }
 }
 #hidden_sync_btn { display: none !important; }
 #tuning_toggle_btn { display: none !important; }
+
+/* SPINNER CSS */
+.loading-box {
+    display: flex; align-items: center; justify-content: center; gap: 12px;
+    padding: 30px; background: #fff; border: 1px solid #E0DFDC;
+    border-radius: 12px; margin-top: 4px; color: #0A66C2; font-weight: 600;
+}
+.spinner {
+    width: 20px; height: 20px; border: 3px solid #EBF3FB;
+    border-top: 3px solid #0A66C2; border-radius: 50%;
+    animation: spin 1s linear infinite;
+}
+@keyframes spin { 0% { transform: rotate(0deg); } 100% { transform: rotate(360deg); } }
 """
 
 # ── UI ─────────────────────────────────────────────────────────────────────────
@@ -696,7 +755,7 @@ with gr.Blocks(title="LinkedIn Translator") as demo:
 
     gr.HTML("""
         <div class="li-footer">
-            <span>🧠 Llama 3.3 70B Instruct</span>
+            <span>🧠 Llama 4 (17B) & Llama 3.3 (70B)</span>
             <span>🔄 Bidirektional</span>
             <span>🎯 Corporate Nonsense Score</span>
             <span>✨ AI Tuning</span>
@@ -746,7 +805,7 @@ with gr.Blocks(title="LinkedIn Translator") as demo:
     )
 
     tuning_btn.click(
-        fn=generate_tuned_post,
+        fn=generate_tuned_post_with_loader,
         inputs=[input_box, slider_ton, slider_substanz, slider_laenge, dd_zielgruppe, dd_cta],
         outputs=[tuning_out]
     )