Update app.py

app.py

@@ -1,93 +1,50 @@
-import gradio as gr  # gradio makes it super easy to build a web UI
-import torch  # torch is used to run the model and handle tensors
-from transformers import BertTokenizer, BertForSequenceClassification  # for loading our fine-tuned BERT model and tokenizer
-import zipfile  # we use this to unzip the uploaded model
-import os  # lets us check if the model folder already exists
-
-# check if model folder is already extracted
-if not os.path.exists("fine_tuned_model"):
-    # if not, unzip it
-    with zipfile.ZipFile("fine_tuned_model.zip", 'r') as zip_ref:
-        zip_ref.extractall("fine_tuned_model")
-
-# path to our model directory
-model_path = "./fine_tuned_model"
-
-# load the tokenizer and model from the directory
-tokenizer = BertTokenizer.from_pretrained(model_path)  # tokenizer breaks text into model-friendly tokens
-model = BertForSequenceClassification.from_pretrained(model_path)  # load the actual fine-tuned BERT model
-model.eval()  # set it to eval mode so it doesn’t try to learn during predictions
-
-# define labels in case we want to refer back to them
-label_map = {
-    0: "Unbiased",
-    1: "Biased"
-}
-
-# the main function that runs when user submits text
 def detect_bias(text):
-    # convert user input into tensors using the tokenizer
     inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
 
-    # disable gradient tracking — we’re only doing prediction, not training
     with torch.no_grad():
-        outputs = model(**inputs)  # pass inputs through the model
-        logits = outputs.logits  # raw prediction scores
-        probs = torch.softmax(logits, dim=1).squeeze()  # turn scores into probabilities
-        pred_label = torch.argmax(probs).item()  # get the predicted label (0 or 1)
-        confidence = round(probs[pred_label].item(), 2)  # grab the confidence score of that prediction
-
-    # classify the result based on thresholds
-    if confidence > 0.75 and pred_label == 1:
-        final_label = "Biased"
-        explanation = (
-            "⚠️ This text is likely biased. The model is highly confident that it reflects gender stereotypes or role bias."
-        )
-    elif 0.56 < confidence <= 0.75 and pred_label == 1:
-        final_label = "Possibly Biased"
-        explanation = (
-            "🤔 This text might contain some gender bias, but the model is not entirely sure. Review it carefully."
-        )
-    elif confidence <= 0.56:
-        final_label = "Unbiased"
-        explanation = (
-            "✅ This text appears neutral with no strong signs of gender bias based on the model's prediction."
-        )
-    else:
-        final_label = label_map[pred_label]
-        explanation = "Prediction complete."
+        outputs = model(**inputs)
+        logits = outputs.logits
+        probs = torch.softmax(logits, dim=1).squeeze()
+        pred_label = torch.argmax(probs).item()
+        confidence = round(probs[pred_label].item(), 2)
+
+    # logic checks both label AND confidence
+    if pred_label == 1:  # predicted as biased
+        if confidence > 0.75:
+            final_label = "Biased"
+            explanation = (
+                "⚠️ This text is likely biased. The model is highly confident that it reflects gender stereotypes or role bias."
+            )
+        elif 0.5 <= confidence <= 0.75:
+            final_label = "Possibly Biased"
+            explanation = (
+                "🤔 This text might contain some gender bias, but the model is not entirely sure. Review it carefully."
+            )
+        else:
+            final_label = "Uncertain"
+            explanation = (
+                "😐 The model predicted 'biased' but with very low confidence. The result may not be reliable."
+            )
+
+    elif pred_label == 0:  # predicted as unbiased
+        if confidence > 0.75:
+            final_label = "Unbiased"
+            explanation = (
+                "✅ This text appears neutral with no strong signs of gender bias based on the model's understanding."
+            )
+        elif 0.5 <= confidence <= 0.75:
+            final_label = "Possibly Unbiased"
+            explanation = (
+                "🤔 This text seems unbiased, but the model isn't highly confident. It may still be worth reviewing."
+            )
+        else:
+            final_label = "Uncertain"
+            explanation = (
+                "😐 The model predicted 'unbiased' but with low confidence. The result is unclear."
+            )
 
-    # send the results back to the UI
     return {
         "Bias Classification": final_label,
         "Confidence Score": confidence,
         "Explanation": explanation
-    }
-
-# build the Gradio web interface
-with gr.Blocks() as demo:
-    # title and description at the top
-    gr.Markdown("## Bias Bin – Fine-Tuned BERT Version by Aryan, Gowtham & Manoj")
-    gr.Markdown("This tool detects **gender bias** in narrative text using a BERT model fine-tuned on custom counterfactual data.")
-
-    # text input box for user
-    text_input = gr.Textbox(
-        label="Enter Narrative Text",
-        lines=4,
-        placeholder="E.g., 'The woman stayed at home while the man went to work.'"
-    )
-
-    # button to trigger prediction
-    submit_btn = gr.Button("Detect Bias")
-
-    # box to show the output (bias label + confidence + explanation)
-    output = gr.JSON(label="Prediction Output")
-
-    # link the button to the function
-    submit_btn.click(fn=detect_bias, inputs=text_input, outputs=output)
-
-    # show a simple disclaimer at the bottom for transparency
-    gr.Markdown("⚠️ **Disclaimer:** This model is trained on a small, augmented dataset and may not always be accurate. Interpret results carefully and consider human review where needed.")
-
-# launch the app (runs on HF Spaces)
-demo.launch()
+    }