Update app.py

app.py

@@ -84,81 +84,60 @@ html, body, [class*="css"] { font-family: 'Inter', sans-serif; }
 """, unsafe_allow_html=True)
 
 # ─────────────────────── CONSTANTS ───────────────────────────────────────────
-MODEL_PATH = "fine_tuned_roberta_model"
-BASE_MODEL  = "xlm-roberta-base"
-MAX_LEN     = 256
-DEVICE      = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+FINETUNED_REPO = "PrachiSandipkumar/Taana-Tracker-Sarcasm"
+BASE_MODEL     = "xlm-roberta-base"
+MAX_LEN        = 256
+DEVICE         = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 
 # ─────────────────────── MODEL LOADING ───────────────────────────────────────
 @st.cache_resource(show_spinner=False)
 def load_model():
+    """   
+    Loads tokenizer from xlm-roberta-base and fine-tuned weights 
+    from the dedicated Hugging Face model repository.
     """
-    Mirrors the notebook exactly:
-    - Tokenizer: xlm-roberta-base  (notebook Cell 22)
-    - Model:     fine_tuned_roberta_model folder (notebook Cell 45)
-
-    The notebook only saves the MODEL (model.save_pretrained), not the tokenizer.
-    So we always load the tokenizer from xlm-roberta-base.
-    To also save the tokenizer, add to your notebook after training:
-        tokenizer.save_pretrained("/content/fine_tuned_roberta_model")
-    """
+    
     load_log = []
+    source = "base-NOT-fine-tuned"
 
-    # Tokenizer ➜ notebook Cell 22 always uses xlm-roberta-base
-    tokenizer_in_folder = (
-        Path(MODEL_PATH).exists() and
-        (Path(MODEL_PATH) / "tokenizer.json").exists()
-    )
-    if tokenizer_in_folder:
-        tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)
-        load_log.append("Tokenizer: loaded from saved model folder")
-    else:
+    # 1. Tokenizer ➔ Always pull from xlm-roberta-base as per Cell 22
+    try:
         tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
-        load_log.append("Tokenizer: loaded from xlm-roberta-base (tokenizer not in model folder)")
-
-    # Model ➜     fine-tuned weights from notebook Cell 45
-    weights_exist = Path(MODEL_PATH).exists() and (
-        (Path(MODEL_PATH) / "model.safetensors").exists() or
-        (Path(MODEL_PATH) / "pytorch_model.bin").exists()
-    )
-
-    model  = None
-    source = ""
-
-    if weights_exist:
-        try:
-            # Initialize base model first
-            model = AutoModelForSequenceClassification.from_pretrained(BASE_MODEL, num_labels=2)
-            
-            # Resize embeddings to match tokenizer vocab size
-            tokenizer_vocab_size = len(tokenizer)
-            model_vocab_size = model.get_input_embeddings().weight.shape[0]
-            
-            if tokenizer_vocab_size != model_vocab_size:
-                load_log.append(f"Resizing model embeddings: {model_vocab_size} → {tokenizer_vocab_size}")
-                model.resize_token_embeddings(tokenizer_vocab_size)
-            
-            # Load the fine-tuned weights
-            if (Path(MODEL_PATH) / "model.safetensors").exists():
-                from safetensors.torch import load_file
-                state_dict = load_file(Path(MODEL_PATH) / "model.safetensors")
-                model.load_state_dict(state_dict, strict=False)
-            else:
-                model = AutoModelForSequenceClassification.from_pretrained(MODEL_PATH, num_labels=2)
+        load_log.append("Tokenizer: Successfully loaded from xlm-roberta-base")
+    except Exception as e:
+        st.error(f"Failed to load base tokenizer: {e}")
+        raise e
+
+    # 2. Model ➔ Pull configuration and weights from your fine-tuned repo
+    try:
+        # Load model with weights from your fine-tuned repository
+        model = AutoModelForSequenceClassification.from_pretrained(
+            FINETUNED_REPO, 
+            num_labels=2
+        )
+        
+        # Check and handle vocab size alignment safely
+        tokenizer_vocab_size = len(tokenizer)
+        model_vocab_size = model.get_input_embeddings().weight.shape[0]
+        
+        if tokenizer_vocab_size != model_vocab_size:
+            load_log.append(f"Resizing embeddings: {model_vocab_size} → {tokenizer_vocab_size}")
+            model.resize_token_embeddings(tokenizer_vocab_size)
             
-            source = "fine-tuned"
-            load_log.append(f"Model: fine-tuned weights loaded from '{MODEL_PATH}'")
-        except Exception as e:
-            load_log.append(f"Model: fine-tuned load FAILED ➜ {e}")
-
-    if model is None:
-        model  = AutoModelForSequenceClassification.from_pretrained(BASE_MODEL, num_labels=2)
+        source = "fine-tuned"
+        load_log.append(f"Model: Fine-tuned weights successfully loaded from HF Hub ({FINETUNED_REPO})")
+        
+    except Exception as e:
+        # Fallback safeguard back to base if repo files fail to initialize
+        load_log.append(f"Model: Fine-tuned load FAILED ➔ {e}")
+        load_log.append("Model: WARNING ➔ Falling back to base xlm-roberta-base.")
+        model = AutoModelForSequenceClassification.from_pretrained(BASE_MODEL, num_labels=2)
         source = "base-NOT-fine-tuned"
-        load_log.append("Model: WARNING ➜ using base model. Fine-tuned weights not found!")
 
     model.to(DEVICE)
     model.eval()
     return model, tokenizer, source, load_log
+   
 
 
 # ─────────────────────── NOTEBOOK FUNCTIONS ──────────────────────────────────