app.py

import gradio as gr
from apscheduler.schedulers.background import BackgroundScheduler
from huggingface_hub import snapshot_download
import pandas as pd
import matplotlib.pyplot as plt

# Dataset paths
LEADERBOARD_PATH = "hf://datasets/alibayram/yapay_zeka_turkce_mmlu_liderlik_tablosu/data/train-00000-of-00001.parquet"
RESPONSES_PATH = "hf://datasets/alibayram/yapay_zeka_turkce_mmlu_model_cevaplari/data/train-00000-of-00001.parquet"
SECTION_RESULTS_PATH = "hf://datasets/alibayram/yapay_zeka_turkce_mmlu_bolum_sonuclari/data/train-00000-of-00001.parquet"

# Load datasets
try:
    leaderboard_data = pd.read_parquet(LEADERBOARD_PATH)
    model_responses_data = pd.read_parquet(RESPONSES_PATH)
    section_results_data = pd.read_parquet(SECTION_RESULTS_PATH)
except Exception as e:
    print(f"Error loading datasets: {e}")
    raise

# Helper functions
def filter_leaderboard(family=None, quantization_level=None):
    df = leaderboard_data.copy()
    if family:
        df = df[df["family"] == family]
    if quantization_level:
        df = df[df["quantization_level"] == quantization_level]
    return df

def search_responses(query, model):
    filtered = model_responses_data[model_responses_data["bolum"].str.contains(query, case=False)]
    selected_columns = ["bolum", "soru", "cevap", model + "_cevap"]
    return filtered[selected_columns]

def plot_section_results():
    fig, ax = plt.subplots(figsize=(10, 6))
    avg_scores = section_results_data.mean(numeric_only=True)
    avg_scores.plot(kind="bar", ax=ax)
    ax.set_title("Average Section-Wise Performance")
    ax.set_ylabel("Accuracy (%)")
    ax.set_xlabel("Sections")
    return fig  # Return the figure object

def add_new_model(model_name, base_model, revision, precision, weight_type, model_type):
    # Simulated model submission logic
    return f"Model '{model_name}' submitted successfully!"

# Gradio app structure
with gr.Blocks(css=".container { max-width: 1200px; margin: auto; }") as app:
    gr.HTML("<h1>🏆 Turkish MMLU Leaderboard</h1>")
    gr.Markdown("Explore, evaluate, and compare AI model performance.")

    with gr.Tabs() as tabs:
        # Leaderboard Tab
        with gr.TabItem("Leaderboard"):
            family_filter = gr.Dropdown(
                choices=leaderboard_data["family"].unique().tolist(), label="Filter by Family", multiselect=False
            )
            quantization_filter = gr.Dropdown(
                choices=leaderboard_data["quantization_level"].unique().tolist(), label="Filter by Quantization Level"
            )
            leaderboard_table = gr.DataFrame(leaderboard_data)
            gr.Button("Apply Filters").click(
                filter_leaderboard, inputs=[family_filter, quantization_filter], outputs=leaderboard_table
            )

        # Model Responses Tab
        with gr.TabItem("Model Responses"):
            model_dropdown = gr.Dropdown(
                choices=leaderboard_data["model"].unique().tolist(), label="Select Model"
            )
            query_input = gr.Textbox(label="Search Query")
            responses_table = gr.DataFrame()
            gr.Button("Search").click(
                search_responses, inputs=[query_input, model_dropdown], outputs=responses_table
            )

        # Section Results Tab
        with gr.TabItem("Section Results"):
            gr.Plot(plot_section_results)
            gr.DataFrame(section_results_data)

        # Submit Model Tab
        with gr.TabItem("Submit Model"):
            gr.Markdown("### Submit Your Model for Evaluation")
            model_name = gr.Textbox(label="Model Name")
            base_model = gr.Textbox(label="Base Model")
            revision = gr.Textbox(label="Revision", placeholder="main")
            precision = gr.Dropdown(
                choices=["float16", "int8", "bfloat16", "float32"], label="Precision", value="float16"
            )
            weight_type = gr.Dropdown(
                choices=["Original", "Delta", "Adapter"], label="Weight Type", value="Original"
            )
            model_type = gr.Dropdown(
                choices=["Transformer", "RNN", "GPT", "Other"], label="Model Type", value="Transformer"
            )
            submit_button = gr.Button("Submit")
            submission_output = gr.Markdown()
            submit_button.click(
                add_new_model,
                inputs=[model_name, base_model, revision, precision, weight_type, model_type],
                outputs=submission_output,
            )

# Scheduler for refreshing datasets
scheduler = BackgroundScheduler()
scheduler.add_job(
    lambda: snapshot_download(repo_id="alibayram", repo_type="dataset", local_dir="cache"), 
    "interval", seconds=1800
)
scheduler.start()

# Launch app
app.queue(default_concurrency_limit=40).launch()