Update app.py

app.py

@@ -1,44 +1,36 @@
-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
+import gradio as gr  # used to build the web interface
+import torch  # used to run the model and handle predictions
+from transformers import BertTokenizer, BertForSequenceClassification  # to load our trained model and tokenizer
+import zipfile  # for extracting the uploaded model
+import os  # to check if folder already exists
 
-# check if model folder is already extracted
+# check if the fine-tuned 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
+# load tokenizer and model
 model_path = "./fine_tuned_model"
+tokenizer = BertTokenizer.from_pretrained(model_path)
+model = BertForSequenceClassification.from_pretrained(model_path)
+model.eval()  # set model to evaluation mode (important for inference)
 
-# 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 just for reference (not used directly in decision now)
-label_map = {
-    0: "Unbiased",
-    1: "Biased"
-}
-
-# the main function that runs when user submits text
+# this function will be triggered when user submits a sentence
 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
+        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)
+
+        # flip label logic because model predictions seem inverted
+        pred_label = 1 - pred_label  # flip 0<->1
 
-    # proper logic: evaluate both label and confidence
-    if pred_label == 1:  # model predicts "biased"
+    # prediction and explanation logic based on flipped label and confidence
+    if pred_label == 1:  # now 1 = biased
         if confidence > 0.75:
             final_label = "Biased"
             explanation = (
@@ -55,7 +47,7 @@ def detect_bias(text):
                 "😐 The model predicted 'biased' but with low confidence. The result may not be reliable."
             )
 
-    elif pred_label == 0:  # model predicts "unbiased"
+    elif pred_label == 0:  # now 0 = unbiased
         if confidence > 0.75:
             final_label = "Unbiased"
             explanation = (
@@ -72,37 +64,35 @@ def detect_bias(text):
                 "😐 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
+# build the Gradio UI
 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.")
+    gr.Markdown("Detect gender bias in text using a BERT model fine-tuned with counterfactual data.")
 
-    # text input box for user
+    # input box for users
     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
+    # button to submit
     submit_btn = gr.Button("Detect Bias")
 
-    # box to show the output (bias label + confidence + explanation)
+    # output area
     output = gr.JSON(label="Prediction Output")
 
-    # link the button to the function
+    # connect button to 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.")
+    # disclaimer at the bottom
+    gr.Markdown("⚠️ **Disclaimer:** This model is trained on a small, synthetic dataset and may not always be accurate. Results should be interpreted cautiously and reviewed by a human.")
 
-# launch the app (runs on HF Spaces)
+# run the app
 demo.launch()