app.py

import gradio as gr
from gradio_leaderboard import Leaderboard
import pandas as pd
from huggingface_hub import snapshot_download, create_repo
from huggingface_hub.utils import RepositoryNotFoundError
import os

from src.about import (
    INTRODUCTION_TEXT,
    LLM_BENCHMARKS_TEXT,
    TITLE,
)
from src.display.css_html_js import custom_css
from src.display.utils import (
    BENCHMARK_COLS,
    COLS,
    AutoEvalColumn,
    fields,
)
from src.envs import API, EVAL_RESULTS_PATH, RESULTS_REPO, TOKEN, OWNER
from src.populate import get_leaderboard_df
from src.evaluation.dynamic_eval import run_dynamic_perplexity_eval

def init_leaderboard(dataframe):
    if dataframe is None:
        raise ValueError("Leaderboard DataFrame is None.")
    
    print("\n=== Initializing Leaderboard ===", flush=True)
    print(f"DataFrame shape: {dataframe.shape}", flush=True)
    print(f"DataFrame columns: {dataframe.columns.tolist()}", flush=True)
    
    return Leaderboard(
        value=dataframe,
        select_columns=[c.name for c in fields(AutoEvalColumn) if not c.hidden],
        search_columns=[AutoEvalColumn.model.name],
        hide_columns=[c.name for c in fields(AutoEvalColumn) if c.hidden],
        filter_columns=[
            AutoEvalColumn.model_type.name,
            AutoEvalColumn.precision.name,
        ],
    )

def run_perplexity_test(model_name, revision, precision):
    """Run perplexity evaluation on demand."""
    import sys
    import traceback
    import gradio as gr
    
    if not model_name:
        return "Please enter a model name."
    
    try:
        # Use stderr for more reliable logging in HF Spaces
        sys.stderr.write(f"\n=== RUNNING PERPLEXITY TEST ===\n")
        sys.stderr.write(f"Model: {model_name}\n")
        sys.stderr.write(f"Revision: {revision}\n")
        sys.stderr.write(f"Precision: {precision}\n")
        sys.stderr.flush()
        
        success, result = run_dynamic_perplexity_eval(model_name, revision, precision)
        sys.stderr.write(f"Evaluation result - Success: {success}, Result: {result}\n")
        sys.stderr.flush()
        
        if success:
            sys.stderr.write("Evaluation succeeded - results saved to dataset\n")
            sys.stderr.flush()
            
            return f"""✅ **Perplexity evaluation completed successfully!**

**Model**: {model_name}
**Perplexity Score**: {result:.4f}

🎉 **Results have been saved to the dataset.**

📋 **To see your results in the leaderboard:**
1. Click on the **🏅 Leaderboard** tab above
2. Refresh the page (Ctrl+R or Cmd+R) 
3. Your model should now appear in the rankings!

💡 **Note**: Due to technical limitations with the leaderboard component, results cannot be updated dynamically. The refresh is necessary to see the latest rankings."""
        else:
            return f"❌ **Evaluation failed**: {result}"
            
    except Exception as e:
        error_msg = str(e)
        traceback_str = traceback.format_exc()
        sys.stderr.write(f"Critical error in run_perplexity_test: {error_msg}\n")
        sys.stderr.write(f"Traceback: {traceback_str}\n")
        sys.stderr.flush()
        return f"❌ **Critical error**: {error_msg}"

# Initialize results repository and directory
try:
    # Try to download existing repository
    try:
        snapshot_download(
            repo_id=RESULTS_REPO,
            local_dir=EVAL_RESULTS_PATH,
            repo_type="dataset",
            tqdm_class=None,
            etag_timeout=30,
            token=TOKEN
        )
    except RepositoryNotFoundError:
        # Create the repository if it doesn't exist
        print(f"Creating new results repository: {RESULTS_REPO}")
        create_repo(
            repo_id=RESULTS_REPO,
            repo_type="dataset",
            private=False,
            token=TOKEN
        )
        # Create local directory
        os.makedirs(EVAL_RESULTS_PATH, exist_ok=True)
except Exception as e:
    print(f"Error initializing results: {e}")
    # Ensure local directory exists even if repo operations fail
    os.makedirs(EVAL_RESULTS_PATH, exist_ok=True)

# Get initial leaderboard data
LEADERBOARD_DF = get_leaderboard_df(EVAL_RESULTS_PATH, COLS, BENCHMARK_COLS)

# Create the Gradio interface
demo = gr.Blocks(css=custom_css)
with demo:
    gr.HTML(TITLE)
    gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")

    with gr.Tabs(elem_classes="tab-buttons") as tabs:
        with gr.TabItem("🏅 Leaderboard", elem_id="leaderboard-tab", id=0):
            leaderboard = init_leaderboard(LEADERBOARD_DF)

        with gr.TabItem("📝 About", elem_id="about-tab", id=1):
            gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")

        with gr.TabItem("🧪 Test Model", elem_id="test-model-tab", id=2):
            gr.Markdown("## Run Perplexity Test\n\nTest any Hugging Face model for perplexity evaluation.")
            
            with gr.Row():
                with gr.Column():
                    model_name = gr.Textbox(label="Model name", placeholder="openai-community/gpt2")
                    revision = gr.Textbox(label="Revision", placeholder="main", value="main")
                    precision = gr.Dropdown(
                        choices=["float16", "bfloat16"],
                        label="Precision",
                        value="float16"
                    )
                    debug_mode = gr.Checkbox(label="Enable debug mode (more verbose logging)", value=True)
                
                with gr.Column():
                    test_button = gr.Button("🚀 Run Perplexity Test", variant="primary")
                    result = gr.Markdown()
            
            gr.Markdown("""
            ### Tips:
            - **Check stderr logs** in HF Spaces for detailed debugging information
            - **After evaluation completes**, click the 🏅 Leaderboard tab and refresh the page to see results
            - **Example models to test**: `openai-community/gpt2`, `EleutherAI/gpt-neo-1.3B`, `openai-community/gpt2-large`
            - **Lower perplexity scores = better performance** (better at predicting text)
            
            ### How it works:
            1. Enter a model name from Hugging Face Hub
            2. Click "Run Perplexity Test" 
            3. Wait for evaluation to complete (may take a few minutes for large models)
            4. Go to 🏅 Leaderboard tab and refresh the page to see your results!
            """)
            
            test_button.click(
                run_perplexity_test,
                [model_name, revision, precision],
                [result]
            )

demo.queue(default_concurrency_limit=5).launch()