app.py

import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
import os
from huggingface_hub import login
from toy_dataset_eval import evaluate_toy_dataset
from mmlu_pro_eval_adapted import evaluate_mmlu_pro
import spaces
import pandas as pd
import time  # Added for timing functionality

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

# ---------------------------------------------------------------------------
# 1. Model and tokenizer setup and Loading
# ---------------------------------------------------------------------------
model_name = "mistralai/Mistral-7B-v0.1"
tokenizer = None
model = None
model_loaded = False
    
# ---------------------------------------------------------------------------
# 1. MMLU-Pro Evaluation call
# ---------------------------------------------------------------------------
@spaces.GPU(duration=120)  # Allow up to 2 minutes for full evaluation
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 (1-57)
        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 (1-20 or -1 for all)
        progress (gr.Progress): Progress indicator
    """
    
    # Convert num_subjects to -1 if all_subjects is True
    if all_subjects:
        num_subjects = -1
        
    # Convert num_questions to -1 if all_questions is True
    if all_questions:
        num_questions = -1

    # Run evaluation with timing
    start_time = time.time()  # Start timing
    results = evaluate_mmlu_pro(
        model_name,
        num_subjects=num_subjects,
        num_questions=num_questions,
        num_shots=num_shots, 
    )
    elapsed_time = time.time() - start_time  # Calculate elapsed 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)
    
    # Verify that the overall accuracy is consistent with the total correct/total samples
    assert abs(overall_acc - (total_correct / total_samples)) < 1e-6, \
        "Overall accuracy calculation mismatch detected"
    
    # 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))
    
# ---------------------------------------------------------------------------
# 4. Gradio Interface
# ---------------------------------------------------------------------------
with gr.Blocks() as demo:
    gr.Markdown("# Mistral-7B on MMLU-Pro Evaluation Demo")
    gr.Markdown("""
    This demo evaluates Mistral-7B-v0.1 on the MMLU-Pro Dataset (available here: https://huggingface.co/datasets/TIGER-Lab/MMLU-Pro).
    """)

    # MMLU Evaluation Section
    gr.Markdown("### MMLU Evaluation")
    
    with gr.Row():
        all_subjects_checkbox = gr.Checkbox(
            label="Evaluate All Subjects",
            value=False,  # Default is unchecked
            info="When checked, evaluates all 14 MMLU-Pro subjects"
        )
        num_subjects_slider = gr.Slider(
            minimum=1,
            maximum=14,
            value=14,  # Default is all subjects
            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,  # Default is 5 few-shot examples
            step=1,
            label="Number of Few-shot Examples",
            info="Number of examples to use for few-shot learning (0-5). They will be loaded in alphabetical order."
        )
    
    with gr.Row():
        all_questions_checkbox = gr.Checkbox(
            label="Evaluate All Questions",
            value=False,  # Default is unchecked
            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")
    
    # Move the slider into the container for easier visibility toggling
    with questions_container:
        num_questions_slider = gr.Slider(
            minimum=1,
            maximum=40,
            value=20,  # Default is 10 questions
            step=1,
            label="Questions per Subject",
            info="Choose a subset of questions (1-40) per subject. They will be loaded in order of question_id for reproducibility. ",
            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 MMLU-Pro Evaluation", variant="stop", visible=False)
        results_output = gr.Markdown(label="Evaluation Results")
        
    with gr.Row():
        results_table = gr.DataFrame(interactive=True, label="Detailed Results (Sortable)", visible=True)
        
    # Update num_subjects_slider interactivity based on all_subjects checkbox
    def update_subjects_slider(checked):
        if checked:
            return gr.update(value=14, interactive=False)
        else:
            return gr.update(interactive=True)
    
    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 disable_ui_for_evaluation():
        return [
            gr.update(interactive=False, info="MMLU Evaluation currently in progress"),  # all_subjects_checkbox
            gr.update(interactive=False, info="MMLU Evaluation currently in progress"),  # num_subjects_slider
            gr.update(interactive=False, info="MMLU Evaluation currently in progress"),  # num_shots_slider
            gr.update(interactive=False, info="MMLU Evaluation currently in progress"),  # all_questions_checkbox
            gr.update(interactive=False, info="MMLU Evaluation currently in progress"),  # num_questions_slider
            gr.update(interactive=False),  # eval_mmlu_button
            gr.update(visible=True)   # cancel_mmlu_button
        ]
    
    # Function to handle cancel button click
    def cancel_evaluation():
        # This doesn't actually cancel the GPU job (which would require more backend support)
        # But it does reset the UI state to be interactive again
        return [
            gr.update(interactive=True, info="When checked, evaluates all 14 MMLU-Pro subjects"),  # all_subjects_checkbox
            gr.update(interactive=True, info="Number of subjects to evaluate (1-14). They will be loaded in alphabetical order."),  # num_subjects_slider
            gr.update(interactive=True, info="Number of examples to use for few-shot learning (0-5). They will be loaded in alphabetical order."),  # num_shots_slider
            gr.update(interactive=True, info="When checked, evaluates all available questions for each subject"),  # all_questions_checkbox
            gr.update(interactive=True, info="Choose a subset of questions (1-40) per subject. They will be loaded in order of question_id for reproducibility."),  # num_questions_slider
            gr.update(interactive=True),  # eval_mmlu_button
            gr.update(visible=False),  # cancel_mmlu_button
            "⚠️ Evaluation canceled by user", # results_output
            None  # results_table
        ]
    
    # Connect MMLU evaluation button - now disables UI and shows cancel button
    eval_mmlu_button.click(
        fn=disable_ui_for_evaluation,
        inputs=None,
        outputs=[
            all_subjects_checkbox,
            num_subjects_slider,
            num_shots_slider,
            all_questions_checkbox,
            num_questions_slider,
            eval_mmlu_button,
            cancel_mmlu_button
        ]
    ).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
        ]
    )
    
    # Connect cancel button
    cancel_mmlu_button.click(
        fn=cancel_evaluation,
        inputs=None,
        outputs=[
            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
        ]
    )

demo.launch()