Update app.py

app.py

@@ -4,14 +4,14 @@ import json
 import torch
 
 # Define hyperparameters
-learning_rate = 5e-5  # Slightly lower learning rate
-batch_size = 8  # Smaller batch size to allow for more precise updates
-epochs = 4  # Slightly more training epochs
-max_seq_length = 256  # Smaller sequence length, especially if the majority of your questions and contexts are shorter
-warmup_steps = 200  # Longer warmup phase
-weight_decay = 0.01  # Keep weight decay as it is
-dropout_prob = 0.2  # Slightly higher dropout for regularization
-gradient_clip_value = 1.0  # Keep gradient clip value as it is
+learning_rate = 3e-5
+batch_size = 16
+epochs = 3
+max_seq_length = 512
+warmup_steps = 100
+weight_decay = 0.01
+dropout_prob = 0.1
+gradient_clip_value = 1.0
 
 context_val = ''
 
@@ -23,8 +23,7 @@ context = gr.Textbox(label="Add the Context (Paragraph or texts) for which you w
 
 def q_n_a_fn(context, text):
     QA_input = {'question': text, 'context': context}
-    nlp = pipeline('question-answering', model=q_n_a_model, tokenizer=tokenizer)
-
+    
     # Set the device (CPU or GPU)
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
     q_n_a_model.to(device)
@@ -34,7 +33,7 @@ def q_n_a_fn(context, text):
 
     # Get predictions
     with torch.no_grad():
-        outputs = q_n_a_model(**inputs)
+        outputs = q_n_a_model(**inputs)  # Use q_n_a_model to get model predictions
     
     # Get the predicted answer span indices
     start_idx, end_idx = torch.argmax(outputs.start_logits), torch.argmax(outputs.end_logits)
@@ -47,7 +46,7 @@ def q_n_a_fn(context, text):
     answer_tokens = inputs["input_ids"][0][start_idx : end_idx + 1]
 
     # Decode the answer tokens into a human-readable answer
-    answer = tokenizer.decode(answer_tokens, skip_special_tokens=True)
+    answer = tokenizer.decode(inputs["input_ids"][0][start_idx:end_idx+1], skip_special_tokens=True)
 
     return answer
 
@@ -85,4 +84,4 @@ with gr.Blocks(theme='gradio/soft') as demo:
             gr.Interface(fn=classification_fn, inputs=[context], outputs="text")
 
 if __name__ == "__main__":
-    demo.launch()
+    demo.launch()