Upload app.py

app.py

@@ -31,7 +31,7 @@ def create_comparison_app(file_paths):
         files_left = len(file_paths) - (current_file_index + 1)
         if current_filepath:
             prompts_left = len(results_data[current_filepath]) - (current_prompt_index + 1) if current_prompt_index < len(results_data[current_filepath]) else 0
-            return f"File {current_file_index + 1}/{len(file_paths)}: {os.path.basename(current_filepath)} - {prompts_left + 1} prompts left in this file, {files_left} files remaining."
+            return f"File {current_file_index + 1}/{len(file_paths)} - {prompts_left + 1} prompts left in this file, {files_left} files remaining."
         else:
             return "No file loaded."
 
@@ -77,7 +77,7 @@ def create_comparison_app(file_paths):
                 else:
                     # All files finished, prepare zip archive for download
                     zip_filepath = create_zip_archive(results_data)
-                    return "Comparison finished for all files! All prompts ranked.", "", "", "Comparison finished for all files!", gr.update(visible=True, value=zip_filepath, label="Download All Results") # Final completion, with file download
+                    return "Comparison finished for all files! Please download the results ('Download results' button).", "", "", "Comparison finished for all files!", gr.update(visible=True, value=zip_filepath, label="Download results") # Final completion, with file download
         else:
             # Should not reach here normally, but handle for robustness - in case record_choice is called after file is finished
             if current_file_index < len(file_paths) - 1:
@@ -88,7 +88,7 @@ def create_comparison_app(file_paths):
             else:
                 # All files finished, prepare zip archive for download
                 zip_filepath = create_zip_archive(results_data)
-                return "Comparison finished for all files! All prompts ranked.", "", "", "Comparison finished for all files!", gr.update(visible=True, value=zip_filepath, label="Download All Results")
+                return "Comparison finished for all files! Please download the results ('Download results' button).", "", "", "Comparison finished for all files!", gr.update(visible=True, value=zip_filepath, label="Download results")
 
     def create_zip_archive(results_data):
         """Creates a zip archive of all result files."""
@@ -102,20 +102,21 @@ def create_comparison_app(file_paths):
 
     with gr.Blocks() as iface:
         progress_markdown = gr.Markdown(get_progress_text()) # Progress indication at the top
-        gr.Markdown("# LLM Response Comparison App")
-        prompt_output = gr.Textbox(label="Prompt", lines=3, interactive=False, max_lines=3) # Fixed lines for prompt
+        gr.Markdown("# LLM song lyrics generation ranking")
+        gr.Markdown("There are 5 files (each with 50 prompts) to compare. For each prompt, choose the better lyrics between Model A and Model B. After you complete all files and prompts, you can download the results.")
+        prompt_output = gr.Textbox(label="Lyrics description", lines=3, interactive=False, max_lines=3) # Fixed lines for prompt
 
         with gr.Row(): # Row for side-by-side outputs
             with gr.Column(): # Column for Finetuned Output
-                finetuned_output_box = gr.Textbox(label="Finetuned Model Output", lines=10, interactive=False, max_lines=10) # Fixed lines for finetuned
+                finetuned_output_box = gr.Textbox(label="Model A", lines=10, interactive=False, max_lines=10) # Fixed lines for finetuned
             with gr.Column(): # Column for Base Output
-                base_output_box = gr.Textbox(label="Base Model Output", lines=10, interactive=False, max_lines=10) # Fixed lines for base
+                base_output_box = gr.Textbox(label="Model B", lines=10, interactive=False, max_lines=10) # Fixed lines for base
 
         with gr.Row(): # Row for buttons
-            finetuned_button = gr.Button("Choose Finetuned as Better")
-            base_button = gr.Button("Choose Base as Better")
+            finetuned_button = gr.Button("Model A is better")
+            base_button = gr.Button("Model B is better")
 
-        file_download_output = gr.DownloadButton(label="Download Results", visible=False) # File output component
+        file_download_output = gr.DownloadButton(label="Download results", visible=False) # File output component
 
         def load_initial_file(files):
             if files: