app.py

import gradio as gr
import os
from huggingface_hub import login
from mmlu_pro_eval_adapted import evaluate_mmlu_pro
import spaces
import pandas as pd
import time
import traceback
from dataset_previews import mmlupro_dataset_preview, format_preview_for_display

# Read token and login
hf_token = os.getenv("HF_READ_WRITE_TOKEN")
if hf_token:
    login(hf_token)
else:
    print("⚠️ No HF_READ_WRITE_TOKEN found in environment")

# ---------------------------------------------------------------------------
# 1. Model configuration
# ---------------------------------------------------------------------------
model_name = "mistralai/Mistral-7B-v0.1"

# ---------------------------------------------------------------------------
# 2. MMLU-Pro Evaluation
# ---------------------------------------------------------------------------
@spaces.GPU(duration=240)
def run_mmlu_evaluation(all_subjects, num_subjects, num_shots, all_questions, num_questions, progress=gr.Progress()):
    """
    Runs the MMLU evaluation with the specified parameters.
    
    Args:
        all_subjects (bool): Whether to evaluate all subjects
        num_subjects (int): Number of subjects to evaluate (1-14)
        num_shots (int): Number of few-shot examples (0-5)
        all_questions (bool): Whether to evaluate all questions per subject
        num_questions (int): Number of examples per subject (1-100 or all)
        progress (gr.Progress): Progress indicator
    """
    try:
        # Convert parameters if needed
        if all_subjects:
            num_subjects = -1
            
        if all_questions:
            num_questions = -1

        # Run evaluation with timing
        start_time = time.time()
        results = evaluate_mmlu_pro(
            model_name,
            num_subjects=num_subjects,
            num_questions=num_questions,
            num_shots=num_shots, 
        )
        elapsed_time = time.time() - start_time

        # Format results
        overall_acc = results["overall_accuracy"]
        min_subject, min_acc = results["min_accuracy_subject"]
        max_subject, max_acc = results["max_accuracy_subject"]
        
        # Create DataFrame from results table
        results_df = pd.DataFrame(results["full_accuracy_table"])
        
        # Calculate totals for the overall row
        total_samples = results_df['Num_samples'].sum()
        total_correct = results_df['Num_correct'].sum()
        
        # Create overall row
        overall_row = pd.DataFrame({
            'Subject': ['**Overall**'],
            'Num_samples': [total_samples],
            'Num_correct': [total_correct],
            'Accuracy': [overall_acc]
        })
        
        # Concatenate overall row with results
        results_df = pd.concat([overall_row, results_df], ignore_index=True)
        
        # Format the report
        report = (
            f"### Overall Results\n"
            f"* Overall Accuracy: {overall_acc:.3f}\n"
            f"* Best Performance: {max_subject} ({max_acc:.3f})\n"
            f"* Worst Performance: {min_subject} ({min_acc:.3f})\n"
            f"* Evaluation completed in {elapsed_time:.2f} seconds\n"
        )

        # Return values that re-enable UI components after completion
        return (report, 
                results_df, 
                gr.update(interactive=True), 
                gr.update(visible=False),
                gr.update(interactive=True), 
                gr.update(interactive=True), 
                gr.update(interactive=True), 
                gr.update(interactive=True), 
                gr.update(interactive=True),
                gr.update(visible=True))
                
    except Exception as e:
        # Handle errors gracefully
        error_trace = traceback.format_exc()
        error_message = f"### Error during evaluation\n```\n{error_trace}\n```"
        
        # Re-enable UI components on error
        return (error_message, 
                None,
                gr.update(interactive=True), 
                gr.update(visible=False),
                gr.update(interactive=True), 
                gr.update(interactive=True), 
                gr.update(interactive=True), 
                gr.update(interactive=True), 
                gr.update(interactive=True),
                gr.update(visible=False))

def format_links_with_bullets(links_text):
    """Format links with bullet points for better readability"""
    lines = links_text.split('\n')
    return "• " + "\n• ".join(lines)

# Function to format dataset preview for better display
def enhanced_format_preview_for_display(preview_data):
    """
    Format the preview data with improved readability
    """
    # Create links with bullet points
    links_value = (
        f"Hugging Face: {preview_data['links']['huggingface']}\n"
        f"GitHub: {preview_data['links']['github']}\n"
        f"Paper: {preview_data['links']['paper']}"
    )
    links_formatted = format_links_with_bullets(links_value)
    
    # Create a table format with better column names
    rows = [
        {"Dataset Property": "Dataset Name", "Details": preview_data["dataset_name"]},
        {"Dataset Property": "Evaluation Type", "Details": preview_data["evaluation_type"]},
        {"Dataset Property": "Description", "Details": preview_data["description"]},
        {"Dataset Property": "Links", "Details": links_formatted},
        {"Dataset Property": "Organization", "Details": preview_data["organization"]},
        {"Dataset Property": "Number of Questions", "Details": preview_data["num_questions"]},
        {"Dataset Property": "Number of Input Tokens", "Details": preview_data["input_tokens"]},
        {"Dataset Property": "Estimated Evaluation Time", "Details": f"{preview_data['evaluation_time']['total_time_minutes']} minutes (for 2 models on A100)"}
    ]
    
    return pd.DataFrame(rows)
    
# ---------------------------------------------------------------------------
# 3. Gradio Interface
# ---------------------------------------------------------------------------
with gr.Blocks() as demo:
    gr.Markdown("# Head-to-Head Model Evaluation Comparator")
    gr.Markdown("""
    This demo evaluates two models (or one model with two different configs) on a benchmark dataset. 
    Available Datasets:[MMLU-Pro](https://huggingface.co/datasets/TIGER-Lab/MMLU-Pro)
    Available Models: [Mistral-7B-v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1)
    """)
    
    # Dataset Selection Section
    gr.Markdown("## (A) Select Dataset for evaluation")
    
    with gr.Row():
        dataset_dropdown = gr.Dropdown(
            choices=["(Select Dataset)", "MMLU-Pro"],
            value="(Select Dataset)",
            label="Dataset",
            info="Select a dataset to perform the Head to Head Evaluation on. Available Datasets: [MMLU-Pro](https://huggingface.co/datasets/TIGER-Lab/MMLU-Pro)"
        )
        preview_toggle = gr.Button("Show Preview", interactive=False, variant="secondary")
    
    # Dataset Preview Container - Initially hidden
    with gr.Column(visible=False) as dataset_preview_container:
        gr.Markdown("## Dataset Preview", elem_id="preview_header")
        preview_output = gr.DataFrame(
            interactive=False,
            wrap=True,
            elem_id="preview_table"
        )
        # Add vertical space after the preview
        gr.Markdown(" ")
        gr.Markdown(" ")
    
    # MMLU Config Container - Initially hidden until dataset is selected
    with gr.Column(visible=False) as mmlu_config_container:
        gr.Markdown("## (B) Select Dataset Configuration Options")
        
        with gr.Row():
            all_subjects_checkbox = gr.Checkbox(
                label="Evaluate All Subjects",
                value=False,
                info="When checked, evaluates all 14 MMLU-Pro subjects"
            )
            num_subjects_slider = gr.Slider(
                minimum=1,
                maximum=14,
                value=14,
                step=1,
                label="Number of Subjects",
                info="Number of subjects to evaluate (1-14). They will be loaded in alphabetical order.",
                interactive=True
            )
        
        with gr.Row():
            num_shots_slider = gr.Slider(
                minimum=0,
                maximum=5,
                value=5,
                step=1,
                label="Number of Few-shot Examples",
                info="Number of examples to use for few-shot learning (0-5)."
            )
        
        with gr.Row():
            all_questions_checkbox = gr.Checkbox(
                label="Evaluate All Questions",
                value=False,
                info="When checked, evaluates all available questions for each subject"
            )
            questions_info_text = gr.Markdown(visible=False, value="**All 12,032 questions across all subjects will be evaluated**")
        
        with gr.Row(elem_id="questions_selection_row"):
            questions_container = gr.Column(scale=1, elem_id="questions_slider_container")
        
        with questions_container:
            num_questions_slider = gr.Slider(
                minimum=1,
                maximum=100,
                value=20,
                step=1,
                label="Questions per Subject",
                info="Choose a subset of questions (1-100) per subject. They will be loaded in order of question_id.",
                interactive=True
            )
        
        with gr.Row():
            with gr.Column(scale=1):
                eval_mmlu_button = gr.Button("Run MMLU-Pro Evaluation", variant="primary", interactive=True)
                cancel_mmlu_button = gr.Button("Cancel Evaluation", variant="stop", visible=False)
    
    # Results Section - Initially hidden
    with gr.Column(visible=False) as results_container:
        results_output = gr.Markdown(label="Evaluation Results")
        
        # Results table - Initially hidden until evaluation completes
        with gr.Column(visible=False) as results_table_container:
            with gr.Row():
                results_table = gr.DataFrame(
                    interactive=True, 
                    label="Detailed Results (Sortable)",
                    visible=True
                )
    
    # Track evaluation state and preview state
    evaluation_state = gr.State({"running": False})
    preview_state = gr.State({"visible": False})
    
    # Function to show/hide configuration based on selected dataset
    def update_interface_based_on_dataset(dataset):
        if dataset == "MMLU-Pro":
            return (
                gr.update(visible=True),  # mmlu_config_container
                gr.update(visible=True),  # results_container
                gr.update(interactive=True)  # preview_toggle
            )
        else:
            return (
                gr.update(visible=False),  # mmlu_config_container
                gr.update(visible=False),  # results_container
                gr.update(interactive=False)  # preview_toggle
            )
    
    # Connect dataset dropdown to show/hide appropriate configuration
    dataset_dropdown.change(
        fn=update_interface_based_on_dataset,
        inputs=[dataset_dropdown],
        outputs=[mmlu_config_container, results_container, preview_toggle]
    )
    
    # Function to toggle dataset preview visibility
    def toggle_preview(state, dataset):
        # Toggle visibility state
        new_visible = not state["visible"]
        state["visible"] = new_visible
        
        # If becoming visible, get the preview data
        if new_visible and dataset == "MMLU-Pro":
            preview_data = mmlupro_dataset_preview()
            formatted_preview = enhanced_format_preview_for_display(preview_data)
            button_text = "Hide Preview"
            return state, gr.update(visible=True), formatted_preview, gr.update(value=button_text)
        elif new_visible:
            # For other datasets (not implemented yet)
            button_text = "Hide Preview"
            return state, gr.update(visible=True), None, gr.update(value=button_text)
        else:
            # Hiding the preview
            button_text = "Show Preview"
            return state, gr.update(visible=False), None, gr.update(value=button_text)
    
    # Connect preview toggle to show/hide dataset information
    preview_toggle.click(
        fn=toggle_preview,
        inputs=[preview_state, dataset_dropdown],
        outputs=[preview_state, dataset_preview_container, preview_output, preview_toggle]
    )
    
    # Update num_subjects_slider interactivity based on all_subjects checkbox
    def update_subjects_slider(checked):
        return gr.update(interactive=not checked)
    
    all_subjects_checkbox.change(
        fn=update_subjects_slider,
        inputs=[all_subjects_checkbox],
        outputs=[num_subjects_slider]
    )
    
    # Update interface based on all_questions checkbox
    def update_questions_interface(checked):
        if checked:
            return gr.update(visible=False), gr.update(visible=True)
        else:
            return gr.update(visible=True), gr.update(visible=False)
    
    all_questions_checkbox.change(
        fn=update_questions_interface,
        inputs=[all_questions_checkbox],
        outputs=[questions_container, questions_info_text]
    )
    
    # Function to disable UI components during evaluation
    def start_evaluation(state):
        if state["running"]:
            return [
                state,
                gr.update(interactive=False),
                gr.update(interactive=False),
                gr.update(interactive=False),
                gr.update(interactive=False),
                gr.update(interactive=False),
                gr.update(interactive=False),
                gr.update(visible=True),
                "Evaluation already in progress. Please wait.",
                None,
                gr.update(visible=False)
            ]
        
        # Update state to running
        state["running"] = True
        
        return [
            state,
            gr.update(interactive=False),  # all_subjects_checkbox
            gr.update(interactive=False),  # num_subjects_slider
            gr.update(interactive=False),  # num_shots_slider
            gr.update(interactive=False),  # all_questions_checkbox
            gr.update(interactive=False),  # num_questions_slider
            gr.update(interactive=False),  # eval_mmlu_button
            gr.update(visible=True),       # cancel_mmlu_button
            "Starting evaluation...",      # results_output
            None,                          # results_table
            gr.update(visible=False)       # results_table_container
        ]
    
    # Function to reset UI after evaluation
    def finish_evaluation(state):
        state["running"] = False
        return state
    
    # Function to handle cancel button click
    def cancel_evaluation(state):
        # Note: This doesn't actually stop the evaluation process
        # It only updates the UI state to appear canceled
        state["running"] = False
        return [
            state,
            gr.update(interactive=True),  # all_subjects_checkbox
            gr.update(interactive=True),  # num_subjects_slider
            gr.update(interactive=True),  # num_shots_slider
            gr.update(interactive=True),  # all_questions_checkbox
            gr.update(interactive=True),  # num_questions_slider
            gr.update(interactive=True),  # eval_mmlu_button
            gr.update(visible=False),     # cancel_mmlu_button
            "⚠️ Evaluation canceled by user (note: backend process may continue running)", # results_output
            None,                         # results_table
            gr.update(visible=False)      # results_table_container
        ]
    
    # Connect MMLU evaluation button with state tracking
    eval_mmlu_button.click(
        fn=start_evaluation,
        inputs=[evaluation_state],
        outputs=[
            evaluation_state,
            all_subjects_checkbox,
            num_subjects_slider,
            num_shots_slider,
            all_questions_checkbox,
            num_questions_slider,
            eval_mmlu_button,
            cancel_mmlu_button,
            results_output,
            results_table,
            results_table_container
        ]
    ).then(
        fn=run_mmlu_evaluation,
        inputs=[
            all_subjects_checkbox,
            num_subjects_slider,
            num_shots_slider,
            all_questions_checkbox,
            num_questions_slider
        ],
        outputs=[
            results_output,
            results_table,
            eval_mmlu_button, 
            cancel_mmlu_button,
            all_subjects_checkbox,
            num_subjects_slider,
            num_shots_slider,
            all_questions_checkbox,
            num_questions_slider,
            results_table_container
        ]
    ).then(
        fn=finish_evaluation,
        inputs=[evaluation_state],
        outputs=[evaluation_state]
    )
    
    # Connect cancel button
    cancel_mmlu_button.click(
        fn=cancel_evaluation,
        inputs=[evaluation_state],
        outputs=[
            evaluation_state,
            all_subjects_checkbox,
            num_subjects_slider,
            num_shots_slider,
            all_questions_checkbox,
            num_questions_slider,
            eval_mmlu_button,
            cancel_mmlu_button,
            results_output,
            results_table,
            results_table_container
        ]
    )

# Add custom CSS for styling
css = """
#preview_header {
    margin-bottom: 10px;
    margin-top: 5px;
}
#preview_table {
    background-color: #f8f9fa;
    border-radius: 8px;
    padding: 10px;
}
"""

demo.launch(css=css)