Update app.py

app.py

@@ -1,80 +1,80 @@
-# app.py
-
-import streamlit as st
-import torch
-import torch.nn as nn
-from transformers import AutoTokenizer, AutoModel
-
-# ✅ Device
-device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-
-# ✅ Load Tokenizer
-tokenizer = AutoTokenizer.from_pretrained("./final_deberta_model")
-
-# ✅ Define Model
-class ScoringModel(nn.Module):
-    def __init__(self, base_model_path="./final_deberta_model", dropout_rate=0.242):
-        super().__init__()
-        self.base = AutoModel.from_pretrained(base_model_path)
-        self.base.gradient_checkpointing_enable()
-        self.dropout1 = nn.Dropout(dropout_rate)
-        self.dropout2 = nn.Dropout(dropout_rate)
-        self.dropout3 = nn.Dropout(dropout_rate)
-        self.classifier = nn.Linear(self.base.config.hidden_size, 1)
-
-    def forward(self, input_ids, attention_mask):
-        hidden = self.base(input_ids=input_ids, attention_mask=attention_mask).last_hidden_state[:, 0]
-        logits = (self.classifier(self.dropout1(hidden)) +
-                  self.classifier(self.dropout2(hidden)) +
-                  self.classifier(self.dropout3(hidden))) / 3
-        return logits
-
-# ✅ Instantiate and Load
-model = ScoringModel()
-model.load_state_dict(torch.load("./final_deberta_model/scoring_model.pt", map_location=device))
-model.to(device)
-model.eval()
-
-# ✅ Prediction function
-def predict(prompt, response_a, response_b):
-    model.eval()
-    with torch.no_grad():
-        text_a = f"Prompt: {prompt} [SEP] {response_a}"
-        text_b = f"Prompt: {prompt} [SEP] {response_b}"
-        
-        encoded_a = tokenizer(text_a, return_tensors='pt', padding="max_length", truncation=True, max_length=186)
-        encoded_b = tokenizer(text_b, return_tensors='pt', padding="max_length", truncation=True, max_length=186)
-        
-        inputs_a = {k: v.to(device) for k, v in encoded_a.items()}
-        inputs_b = {k: v.to(device) for k, v in encoded_b.items()}
-        
-        score_a = model(**inputs_a).squeeze()
-        score_b = model(**inputs_b).squeeze()
-        
-        prob_a = torch.sigmoid(score_a).item()
-        prob_b = torch.sigmoid(score_b).item()
-        
-        return prob_a, prob_b
-
-# ✅ Streamlit App
-st.title("🔍 Fine-Tuned DeBERTa-v3-small: Response Quality Evaluator")
-
-prompt = st.text_area("Enter your prompt:", height=100)
-response_a = st.text_area("Enter Response A:", height=100)
-response_b = st.text_area("Enter Response B:", height=100)
-
-if st.button("Predict Better Response"):
-    if prompt and response_a and response_b:
-        prob_a, prob_b = predict(prompt, response_a, response_b)
-        
-        st.write(f"🔵 **Response A Probability:** {prob_a:.4f}")
-        st.write(f"🟠 **Response B Probability:** {prob_b:.4f}")
-        
-        if prob_b > prob_a:
-            st.success("✅ Model predicts: **Response B** is better!")
-        else:
-            st.success("✅ Model predicts: **Response A** is better!")
-    else:
-        st.warning("⚠️ Please fill in all fields before predicting.")
-
-
+# app.py
+
+import streamlit as st
+import torch
+import torch.nn as nn
+from transformers import AutoModel, PreTrainedTokenizerFast
+
+# ✅ Device
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+# ✅ Load Tokenizer
+tokenizer = PreTrainedTokenizerFast(tokenizer_file="final_deberta_model/tokenizer.json")
+
+# ✅ Define Model
+class ScoringModel(nn.Module):
+    def __init__(self, base_model_name="microsoft/deberta-v3-small", dropout_rate=0.242):
+        super().__init__()
+        self.base = AutoModel.from_pretrained(base_model_name)
+        self.base.gradient_checkpointing_enable()
+        self.dropout1 = nn.Dropout(dropout_rate)
+        self.dropout2 = nn.Dropout(dropout_rate)
+        self.dropout3 = nn.Dropout(dropout_rate)
+        self.classifier = nn.Linear(self.base.config.hidden_size, 1)
+
+    def forward(self, input_ids, attention_mask):
+        hidden = self.base(input_ids=input_ids, attention_mask=attention_mask).last_hidden_state[:, 0]
+        logits = (self.classifier(self.dropout1(hidden)) +
+                  self.classifier(self.dropout2(hidden)) +
+                  self.classifier(self.dropout3(hidden))) / 3
+        return logits
+
+# ✅ Instantiate and Load
+model = ScoringModel()
+model.load_state_dict(torch.load("final_deberta_model/scoring_model.pt", map_location=device))
+model.to(device)
+model.eval()
+
+# ✅ Prediction function
+def predict(prompt, response_a, response_b):
+    model.eval()
+    with torch.no_grad():
+        text_a = f"Prompt: {prompt} [SEP] {response_a}"
+        text_b = f"Prompt: {prompt} [SEP] {response_b}"
+        
+        encoded_a = tokenizer(text_a, return_tensors='pt', padding="max_length", truncation=True, max_length=186)
+        encoded_b = tokenizer(text_b, return_tensors='pt', padding="max_length", truncation=True, max_length=186)
+        
+        inputs_a = {k: v.to(device) for k, v in encoded_a.items()}
+        inputs_b = {k: v.to(device) for k, v in encoded_b.items()}
+        
+        score_a = model(**inputs_a).squeeze()
+        score_b = model(**inputs_b).squeeze()
+        
+        prob_a = torch.sigmoid(score_a).item()
+        prob_b = torch.sigmoid(score_b).item()
+        
+        return prob_a, prob_b
+
+# ✅ Streamlit App
+st.title("🔍 Fine-Tuned DeBERTa-v3-small: Response Quality Evaluator")
+
+prompt = st.text_area("Enter your prompt:", height=100)
+response_a = st.text_area("Enter Response A:", height=100)
+response_b = st.text_area("Enter Response B:", height=100)
+
+if st.button("Predict Better Response"):
+    if prompt and response_a and response_b:
+        prob_a, prob_b = predict(prompt, response_a, response_b)
+        
+        st.write(f"🔵 **Response A Probability:** {prob_a:.4f}")
+        st.write(f"🟠 **Response B Probability:** {prob_b:.4f}")
+        
+        if prob_b > prob_a:
+            st.success("✅ Model predicts: **Response B** is better!")
+        else:
+            st.success("✅ Model predicts: **Response A** is better!")
+    else:
+        st.warning("⚠️ Please fill in all fields before predicting.")
+
+