Update app.py

app.py

@@ -8,10 +8,10 @@ tokenizer = AutoTokenizer.from_pretrained(model_name)
 model = AutoModelForSequenceClassification.from_pretrained(model_name)
 model.eval()
 
-# Function to compute relevance score and dynamically adjust threshold
+# Function to compute relevance score (in logits) and dynamically adjust threshold
 def get_relevance_score_and_excerpt(query, paragraph, threshold_weight):
     if not query.strip() or not paragraph.strip():
-        return "Please provide both a query and a document paragraph.", "", ""
+        return "Please provide both a query and a document paragraph.", ""
 
     # Tokenize the input
     inputs = tokenizer(query, paragraph, return_tensors="pt", truncation=True, padding=True)
@@ -19,13 +19,12 @@ def get_relevance_score_and_excerpt(query, paragraph, threshold_weight):
     with torch.no_grad():
         output = model(**inputs, output_attentions=True)
 
-    # Extract logits and calculate base relevance score
+    # Extract logits (no sigmoid applied)
     logit = output.logits.squeeze().item()
-    base_relevance_score = torch.sigmoid(torch.tensor(logit)).item()
+    base_relevance_score = logit  # Relevance score in logits
 
-    # Calculate dynamic threshold using sigmoid-based formula
-    sigmoid_factor = 1 / (1 + torch.exp(torch.tensor(-5 * (base_relevance_score - 0.5)))).item()
-    dynamic_threshold = max(0.02, threshold_weight * sigmoid_factor)
+    # Dynamically adjust the attention threshold based on user weight (no relevance score influence)
+    dynamic_threshold = max(0.02, threshold_weight)
 
     # Extract attention scores (last layer)
     attention = output.attentions[-1]
@@ -39,12 +38,12 @@ def get_relevance_score_and_excerpt(query, paragraph, threshold_weight):
     para_end_idx = len(inputs["input_ids"][0]) - 1
 
     if para_end_idx <= para_start_idx:
-        return round(base_relevance_score, 4), round(dynamic_threshold, 4), "No relevant tokens extracted."
+        return round(base_relevance_score, 4), "No relevant tokens extracted."
 
     para_attention_scores = attention_scores[para_start_idx:para_end_idx, para_start_idx:para_end_idx].mean(dim=0)
 
     if para_attention_scores.numel() == 0:
-        return round(base_relevance_score, 4), round(dynamic_threshold, 4), "No relevant tokens extracted."
+        return round(base_relevance_score, 4), "No relevant tokens extracted."
 
     # Get indices of relevant tokens above dynamic threshold
     relevant_indices = (para_attention_scores > dynamic_threshold).nonzero(as_tuple=True)[0].tolist()
@@ -59,7 +58,7 @@ def get_relevance_score_and_excerpt(query, paragraph, threshold_weight):
 
     highlighted_text = tokenizer.convert_tokens_to_string(highlighted_text.split())
 
-    return round(base_relevance_score, 4), round(dynamic_threshold, 4), highlighted_text
+    return round(base_relevance_score, 4), highlighted_text
 
 # Define Gradio interface with a slider for threshold adjustment
 interface = gr.Interface(
@@ -67,15 +66,14 @@ interface = gr.Interface(
     inputs=[
         gr.Textbox(label="Query", placeholder="Enter your search query..."),
         gr.Textbox(label="Document Paragraph", placeholder="Enter a paragraph to match..."),
-        gr.Slider(minimum=0.02, maximum=0.5, value=0.1, step=0.01, label="Threshold Weight")
+        gr.Slider(minimum=0.02, maximum=0.5, value=0.1, step=0.01, label="Attention Threshold")
     ],
     outputs=[
-        gr.Textbox(label="Relevance Score"),
-        gr.Textbox(label="Dynamic Threshold"),
+        gr.Textbox(label="Relevance Score (Logits)"),
         gr.HTML(label="Highlighted Document Paragraph")
     ],
     title="Cross-Encoder Attention Highlighting",
-    description="Adjust the threshold weight to influence dynamic token highlighting based on relevance.",
+    description="Adjust the attention threshold to control token highlighting sensitivity.",
     allow_flagging="never",
     live=True
 )