Update app.py

app.py

@@ -2,7 +2,7 @@ import gradio as gr
 import torch
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
-# --- MODEL DATA (Original Detailed Descriptions) ---
+# --- MODEL DATA ---
 MODELS_CONFIG = {
     "Phase 2: Stable (Formal)": {
         "id": "st192011/Maltese-EuroLLM-1.7B-Phase2-Stable",
@@ -30,10 +30,9 @@ MODELS_CONFIG = {
 }
 
 # --- MODEL LOADING (Local CPU) ---
-# Note: Loading two 1.7B models takes ~14GB of RAM. 
-print("Loading models to CPU... Please wait.")
+print("Loading models... this might take a minute.")
 
-# Load Model Phase 2
+# Load Phase 2
 tokenizer_p2 = AutoTokenizer.from_pretrained(MODELS_CONFIG["Phase 2: Stable (Formal)"]["id"])
 model_p2 = AutoModelForCausalLM.from_pretrained(
     MODELS_CONFIG["Phase 2: Stable (Formal)"]["id"], 
@@ -41,7 +40,7 @@ model_p2 = AutoModelForCausalLM.from_pretrained(
     torch_dtype=torch.float32
 )
 
-# Load Model Phase 4
+# Load Phase 4
 tokenizer_p4 = AutoTokenizer.from_pretrained(MODELS_CONFIG["Phase 4: Anchored (Native)"]["id"])
 model_p4 = AutoModelForCausalLM.from_pretrained(
     MODELS_CONFIG["Phase 4: Anchored (Native)"]["id"], 
@@ -53,7 +52,6 @@ def local_translate(model, tokenizer, text, temp):
     if not text.strip():
         return ""
         
-    # Prompt format consistent with training
     prompt = f"### INGLIŻ: {text}\n### MALTI:"
     inputs = tokenizer(prompt, return_tensors="pt").to("cpu")
     
@@ -67,16 +65,20 @@ def local_translate(model, tokenizer, text, temp):
             pad_token_id=tokenizer.eos_token_id
         )
     
-    # skip_special_tokens=True removes the <|endoftext|> and other technical tokens
-    full_text = tokenizer.decode(output_tokens[0], skip_special_tokens=True)
+    # 1. Decode while skipping standard special tokens
+    decoded_text = tokenizer.decode(output_tokens[0], skip_special_tokens=True)
     
-    # Extract only the Maltese translation part (the text after the prompt)
-    if "### MALTI:" in full_text:
-        maltese_text = full_text.split("### MALTI:")[-1].strip()
+    # 2. Extract only the Maltese part
+    if "### MALTI:" in decoded_text:
+        maltese_text = decoded_text.split("### MALTI:")[-1]
     else:
-        maltese_text = full_text.strip()
+        maltese_text = decoded_text
         
-    return maltese_text
+    # 3. CRITICAL: Manual cleaning of the end-of-text string if it still remains
+    # This removes <|endoftext|>, </s>, and any extra whitespace
+    clean_text = maltese_text.replace("<|endoftext|>", "").replace("</s>", "").strip()
+    
+    return clean_text
 
 def translate_logic(text, selected_models, temp):
     out_p2 = "Model not selected."
@@ -103,36 +105,26 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
     
     with gr.Row():
         with gr.Column(scale=2):
-            input_text = gr.Textbox(
-                label="English Source Text", 
-                placeholder="Enter English text here...", 
-                lines=4
-            )
+            input_text = gr.Textbox(label="English Source Text", placeholder="Enter English text...", lines=4)
             model_selector = gr.CheckboxGroup(
                 choices=list(MODELS_CONFIG.keys()), 
                 value=list(MODELS_CONFIG.keys()), 
-                label="Select Models to Compare"
-            )
-            temp_slider = gr.Slider(
-                minimum=0.1, 
-                maximum=1.0, 
-                value=0.1, 
-                step=0.1, 
-                label="Creativity (Temperature)"
+                label="Select Models"
             )
+            temp_slider = gr.Slider(minimum=0.1, maximum=1.0, value=0.1, step=0.1, label="Temperature")
             btn = gr.Button("🚀 Run Translation", variant="primary")
 
     with gr.Row():
         with gr.Column():
             gr.Markdown("### Phase 2: Stable (Formal)")
             p2_out = gr.Textbox(label="Output", interactive=False, lines=5)
-            gr.Markdown(f"**Training Strategy:**\n{MODELS_CONFIG['Phase 2: Stable (Formal)']['description']}")
+            gr.Markdown(f"**Training:** {MODELS_CONFIG['Phase 2: Stable (Formal)']['description']}")
             gr.Markdown(f"**Metrics:** ChrF++: `{MODELS_CONFIG['Phase 2: Stable (Formal)']['chrf']}` | COMET: `{MODELS_CONFIG['Phase 2: Stable (Formal)']['comet']}`")
             
         with gr.Column():
             gr.Markdown("### Phase 4: Anchored (Native)")
             p4_out = gr.Textbox(label="Output", interactive=False, lines=5)
-            gr.Markdown(f"**Training Strategy:**\n{MODELS_CONFIG['Phase 4: Anchored (Native)']['description']}")
+            gr.Markdown(f"**Training:** {MODELS_CONFIG['Phase 4: Anchored (Native)']['description']}")
             gr.Markdown(f"**Metrics:** ChrF++: `{MODELS_CONFIG['Phase 4: Anchored (Native)']['chrf']}` | COMET: `{MODELS_CONFIG['Phase 4: Anchored (Native)']['comet']}`")
 
     gr.Examples(