updsated UI and functionality

app.py

@@ -15,84 +15,69 @@ print("Starting model loading...")
 model = load_model()
 print("Model loaded successfully!")
 
-# Function to generate and evaluate content
-def generate_and_evaluate(preconfigured_prompt):
-    # Step 1: Generate content
-    generation_prompt = [
-        {"role": "user", "content": preconfigured_prompt}
-    ]
-    generated_response = model.create_chat_completion(
-        messages=generation_prompt,
-        max_tokens=256,
-        temperature=1.5
-    )
-    generated_content = generated_response['choices'][0]['message']['content']
-    
-    # Step 2: Evaluate the generated content
+# Function to evaluate two responses
+def evaluate_responses(prompt, response_a, response_b, evaluation_criteria):
+    # Format the evaluation prompt
     evaluation_prompt = [
-        {"role": "system", "content": "You are a strict language evaluator who provides binary assessments of texts."},
-        {"role": "user", "content": f"""Carefully evaluate the generated story:
-Prompt: {preconfigured_prompt}
-Generated response: {generated_content}
-Provide a clear evaluation as follows:
-For each question, write the full question followed by your "Yes" or "No" answer.
-Example format:
-1. Is the story exactly 50 words? - Yes
-2. Does the story contain the letter 'a'? - No
-Now answer these questions:
-1. Is the story exactly 50 words?
-2. Does the story contain the letter 'a'?
-3. Does the story contain the word "alabaster"?
-4. Does the reader understand that the cat's name is Alabaster?
-5. Is the story 100% in English?
-6. Does the text rhyme?"""}
+        {"role": "system", "content": "You are an objective and thorough evaluator of instruction-based responses."},
+        {"role": "user", "content": f"""
+Prompt: {prompt}
+
+Response A: {response_a}
+Response B: {response_b}
+
+Please evaluate both responses based on the following criteria: {evaluation_criteria}
+
+For each criterion, provide a rating of the responses on a scale from 1 to 10, and explain why each response earned that rating. Then, declare a winner (or 'draw' if both are equal).
+"""}
     ]
+    
+    # Generate the evaluation
     evaluation_response = model.create_chat_completion(
         messages=evaluation_prompt,
-        max_tokens=128,
-        temperature=0.2
+        max_tokens=512,
+        temperature=0.5
     )
+    
     evaluation_results = evaluation_response['choices'][0]['message']['content']
     
-    return generated_content, evaluation_results
-
-# Preconfigured prompt
-PRECONFIGURED_PROMPT = """Write a story about the cat Alabaster. It should be exactly 50 words and you are not allowed to use the letter 'a'. The reader must understand that the cat's name is Alabaster. Only replacing the letter 'a' with something like "_" is not enough. The text should rhyme."""
+    return evaluation_results
 
 # Gradio interface
 with gr.Blocks(title="LLM as a Judge") as demo:
     gr.Markdown("## LLM as a Judge 🧐")
-    
-    generate_evaluate_button = gr.Button("Judge the LLM!")
 
-    # Label for the preconfigured prompt
-    gr.Label("Preconfigured prompt:")
-    gr.Label(PRECONFIGURED_PROMPT)
+    # Input fields for the prompt, two responses, and selection of criteria
+    prompt_input = gr.Textbox(label="Enter the Prompt", placeholder="Enter the prompt here...", lines=3)
+    response_a_input = gr.Textbox(label="Response A", placeholder="Enter Response A here...", lines=5)
+    response_b_input = gr.Textbox(label="Response B", placeholder="Enter Response B here...", lines=5)
     
-    generated_output = gr.Textbox(
-        label="Generated Content", 
-        placeholder="The generated content will appear here...", 
-        lines=5, 
-        interactive=False
+    # Dropdown for selecting evaluation criteria
+    criteria_dropdown = gr.Dropdown(
+        label="Select Evaluation Criteria",
+        choices=["Clarity", "Completeness", "Accuracy", "Relevance", "User-Friendliness", "Depth", "Creativity"],
+        value="Clarity",
+        type="value"
     )
     
+    # Button to start the evaluation
+    evaluate_button = gr.Button("Evaluate Responses")
+
+    # Label for displaying the evaluation results
     evaluation_output = gr.Textbox(
         label="Evaluation Results", 
         placeholder="The evaluation results will appear here...", 
-        lines=8, 
+        lines=10, 
         interactive=False
     )
     
-    # Link generation and evaluation
-    generate_evaluate_button.click(
-        fn=generate_and_evaluate,
-        inputs=[gr.State(PRECONFIGURED_PROMPT)],
-        outputs=[generated_output, evaluation_output]
+    # Link evaluation function to the button
+    evaluate_button.click(
+        fn=evaluate_responses,
+        inputs=[prompt_input, response_a_input, response_b_input, criteria_dropdown],
+        outputs=[evaluation_output]
     )
 
 # Launch the app
-demo.launch(
-    server_name="0.0.0.0",
-    server_port=7860,
-    share=False
-)
+if __name__ == "__main__":
+    demo.launch()