app.py

import json
import gzip
import gradio as gr
from gradio_leaderboard import Leaderboard, ColumnFilter, SelectColumns
import pandas as pd
import numpy as np
from apscheduler.schedulers.background import BackgroundScheduler
from huggingface_hub import snapshot_download
from io import StringIO
from typing import Dict, List, Optional
from dataclasses import dataclass, field
from copy import deepcopy

from src.about import (
    CITATION_BUTTON_LABEL,
    CITATION_BUTTON_TEXT,
    EVALUATION_QUEUE_TEXT,
    INTRODUCTION_TEXT,
    LLM_BENCHMARKS_TEXT,
    TITLE,
)
from src.display.css_html_js import custom_css
from src.display.utils import (
    BENCHMARK_COLS,
    BENCHMARK_COLS_MULTIMODAL,
    BENCHMARK_COLS_MIB_SUBGRAPH,
    COLS,
    COLS_MIB_SUBGRAPH,
    COLS_MULTIMODAL,
    EVAL_COLS,
    EVAL_TYPES,
    AutoEvalColumn,
    AutoEvalColumn_mib_subgraph,
    AutoEvalColumn_mib_causalgraph,
    fields,
)
from src.envs import API, EVAL_REQUESTS_PATH, QUEUE_REPO, REPO_ID, TOKEN, RESULTS_REPO_MIB_SUBGRAPH, EVAL_RESULTS_MIB_SUBGRAPH_PATH, RESULTS_REPO_MIB_CAUSALGRAPH, EVAL_RESULTS_MIB_CAUSALGRAPH_PATH
from src.populate import get_evaluation_queue_df, get_leaderboard_df, get_leaderboard_df_mib_subgraph, get_leaderboard_df_mib_causalgraph
from src.submission.submit import add_new_eval


from src.about import TasksMib_Subgraph, TasksMib_Causalgraph

# class SmartSelectColumns(SelectColumns):
#     """
#     Enhanced SelectColumns component with basic filtering functionality.
#     """
#     def __init__(
#         self,
#         benchmark_keywords: Optional[List[str]] = None,
#         model_keywords: Optional[List[str]] = None,
#         initial_selected: Optional[List[str]] = None,
#         **kwargs
#     ):
#         """
#         Initialize SmartSelectColumns with minimal configuration.
        
#         Args:
#             benchmark_keywords: List of benchmark names to filter by
#             model_keywords: List of model names to filter by
#             initial_selected: List of columns to show initially
#         """
#         super().__init__(**kwargs)
#         self.benchmark_keywords = benchmark_keywords or []
#         self.model_keywords = model_keywords or []
#         self.initial_selected = initial_selected or []

#     def get_filtered_groups(self, df: pd.DataFrame) -> Dict[str, List[str]]:
#         """
#         Create column groups based on simple substring matching.
#         """
#         filtered_groups = {}
        
#         # Create benchmark groups
#         for benchmark in self.benchmark_keywords:
#             matching_cols = [
#                 col for col in df.columns 
#                 if benchmark in col.lower()
#             ]
#             if matching_cols:
#                 group_name = f"Benchmark group for {benchmark}"
#                 filtered_groups[group_name] = matching_cols
        
#         # Create model groups
#         for model in self.model_keywords:
#             matching_cols = [
#                 col for col in df.columns 
#                 if model in col.lower()
#             ]
#             if matching_cols:
#                 group_name = f"Model group for {model}"
#                 filtered_groups[group_name] = matching_cols
        
#         return filtered_groups

#     def update(
#         self, 
#         value: Union[pd.DataFrame, Dict[str, List[str]], Any]
#     ) -> Dict:
#         """Update component with new values."""
#         if isinstance(value, pd.DataFrame):
#             choices = list(value.columns)
#             selected = self.initial_selected if self.initial_selected else choices
#             filtered_cols = self.get_filtered_groups(value)
            
#             return {
#                 "choices": choices,
#                 "value": selected,
#                 "filtered_cols": filtered_cols
#             }
        
#         if hasattr(value, '__dataclass_fields__'):
#             field_names = [field.name for field in fields(value)]
#             return {
#                 "choices": field_names,
#                 "value": self.initial_selected if self.initial_selected else field_names
#             }
            
#         return super().update(value)

from gradio_leaderboard import SelectColumns, Leaderboard
import pandas as pd
from typing import List, Dict, Optional
from dataclasses import fields
import math

class SmartSelectColumns(SelectColumns):
    """
    Enhanced SelectColumns component matching exact original parameters.
    """
    def __init__(
        self,
        benchmark_keywords: Optional[List[str]] = None,
        model_keywords: Optional[List[str]] = None,
        initial_selected: Optional[List[str]] = None,
        label: Optional[str] = None,
        show_label: bool = True,
        info: Optional[str] = None,
        allow: bool = True
    ):
        # Match exact parameters from working SelectColumns
        super().__init__(
            default_selection=initial_selected or [],
            cant_deselect=[],
            allow=allow,
            label=label,
            show_label=show_label,
            info=info
        )
        
        self.benchmark_keywords = benchmark_keywords or []
        self.model_keywords = model_keywords or []
        
        # Store groups for later use
        self._groups = {}
    
    def get_filtered_groups(self, columns: List[str]) -> Dict[str, List[str]]:
        """Get column groups based on keywords."""
        filtered_groups = {}
        
        # Add benchmark groups
        for benchmark in self.benchmark_keywords:
            matching_cols = [
                col for col in columns 
                if benchmark in col.lower()
            ]
            if matching_cols:
                filtered_groups[f"Benchmark group for {benchmark}"] = matching_cols
        
        # Add model groups
        for model in self.model_keywords:
            matching_cols = [
                col for col in columns 
                if model in col.lower()
            ]
            if matching_cols:
                filtered_groups[f"Model group for {model}"] = matching_cols
                
        self._groups = filtered_groups
        return filtered_groups


import re
@dataclass
class SubstringSelectColumns(SelectColumns):
    """
    Extends SelectColumns to support filtering columns by predefined substrings.
    When a substring is selected, all columns containing that substring will be selected.
    """
    substring_groups: Dict[str, List[str]] = field(default_factory=dict)
    selected_substrings: List[str] = field(default_factory=list)
    
    def __post_init__(self):
        # Ensure default_selection is a list
        if self.default_selection is None:
            self.default_selection = []
            
        # Build reverse mapping of column to substrings
        self.column_to_substrings = {}
        for substring, patterns in self.substring_groups.items():
            for pattern in patterns:
                # Convert glob-style patterns to regex
                regex = re.compile(pattern.replace('*', '.*'))
                # Find matching columns in default_selection
                for col in self.default_selection:
                    if regex.search(col):
                        if col not in self.column_to_substrings:
                            self.column_to_substrings[col] = []
                        self.column_to_substrings[col].append(substring)
        
        # Apply initial substring selections
        if self.selected_substrings:
            self.update_selection_from_substrings()
    
    def update_selection_from_substrings(self) -> List[str]:
        """
        Updates the column selection based on selected substrings.
        Returns the new list of selected columns.
        """
        selected_columns = self.cant_deselect.copy()
        
        # If no substrings selected, show all columns
        if not self.selected_substrings:
            selected_columns.extend([
                col for col in self.default_selection 
                if col not in self.cant_deselect
            ])
            return selected_columns
        
        # Add columns that match any selected substring
        for col, substrings in self.column_to_substrings.items():
            if any(s in self.selected_substrings for s in substrings):
                if col not in selected_columns:
                    selected_columns.append(col)
                    
        return selected_columns





def restart_space():
    API.restart_space(repo_id=REPO_ID)



### Space initialisation
try:
    # print(EVAL_REQUESTS_PATH)
    snapshot_download(
        repo_id=QUEUE_REPO, local_dir=EVAL_REQUESTS_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30, token=TOKEN
    )
except Exception:
    restart_space()


try:
    # print(RESULTS_REPO_MIB_SUBGRAPH)
    snapshot_download(
        repo_id=RESULTS_REPO_MIB_SUBGRAPH, local_dir=EVAL_RESULTS_MIB_SUBGRAPH_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30, token=TOKEN
    )
except Exception:
    restart_space()


try:
    # print(RESULTS_REPO_MIB_CAUSALGRAPH)
    snapshot_download(
        repo_id=RESULTS_REPO_MIB_CAUSALGRAPH, local_dir=EVAL_RESULTS_MIB_CAUSALGRAPH_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30, token=TOKEN
    )
except Exception:
    restart_space()

def _sigmoid(x):
    try:
        return 1 / (1 + math.exp(-2 * (x-1)))
    except:
        return "-"

LEADERBOARD_DF_MIB_SUBGRAPH_FPL = get_leaderboard_df_mib_subgraph(EVAL_RESULTS_MIB_SUBGRAPH_PATH, EVAL_REQUESTS_PATH, COLS_MIB_SUBGRAPH, BENCHMARK_COLS_MIB_SUBGRAPH)
LEADERBOARD_DF_MIB_SUBGRAPH_FEQ = get_leaderboard_df_mib_subgraph(EVAL_RESULTS_MIB_SUBGRAPH_PATH, EVAL_REQUESTS_PATH, COLS_MIB_SUBGRAPH, BENCHMARK_COLS_MIB_SUBGRAPH,
                                                                  metric_type="F=")

# LEADERBOARD_DF_MIB_CAUSALGRAPH = get_leaderboard_df_mib_causalgraph(EVAL_RESULTS_MIB_CAUSALGRAPH_PATH, EVAL_REQUESTS_PATH, COLS_MIB_CAUSALGRAPH, BENCHMARK_COLS_MIB_CAUSALGRAPH)
# In app.py, modify the LEADERBOARD initialization
LEADERBOARD_DF_MIB_CAUSALGRAPH_DETAILED, LEADERBOARD_DF_MIB_CAUSALGRAPH_AGGREGATED, LEADERBOARD_DF_MIB_CAUSALGRAPH_AVERAGED = get_leaderboard_df_mib_causalgraph(
    EVAL_RESULTS_MIB_CAUSALGRAPH_PATH, 
    EVAL_REQUESTS_PATH
)


# LEADERBOARD_DF = get_leaderboard_df(EVAL_RESULTS_PATH, EVAL_REQUESTS_PATH, COLS, BENCHMARK_COLS)
# LEADERBOARD_DF_MULTIMODAL = get_leaderboard_df(EVAL_RESULTS_PATH, EVAL_REQUESTS_PATH, COLS_MULTIMODAL, BENCHMARK_COLS_MULTIMODAL)

(
    finished_eval_queue_df,
    running_eval_queue_df,
    pending_eval_queue_df,
) = get_evaluation_queue_df(EVAL_REQUESTS_PATH, EVAL_COLS)





def init_leaderboard_mib_subgraph(dataframe, track):
    """Initialize the subgraph leaderboard with display names for better readability."""
    if dataframe is None or dataframe.empty:
        raise ValueError("Leaderboard DataFrame is empty or None.")

    print("\nDebugging DataFrame columns:", dataframe.columns.tolist())
    
    # First, create our display name mapping
    # This is like creating a translation dictionary between internal names and display names
    model_name_mapping = {
        "qwen2_5": "Qwen-2.5",
        "gpt2": "GPT-2",
        "gemma2": "Gemma-2",
        "llama3": "Llama-3.1"
    }

    benchmark_mapping = {
        "ioi": "IOI",
        "mcqa": "MCQA",
        "arithmetic_addition": "Arithmetic (+)",
        "arithmetic_subtraction": "Arithmetic (-)",
        "arc_easy": "ARC (Easy)",
        "arc_challenge": "ARC (Challenge)"
    }

    display_mapping = {}
    for task in TasksMib_Subgraph:
        for model in task.value.models:
            field_name = f"{task.value.benchmark}_{model}"
            display_name = f"{benchmark_mapping[task.value.benchmark]} - {model_name_mapping[model]}"
            display_mapping[field_name] = display_name
    

    # Now when creating benchmark groups, we'll use display names
    benchmark_groups = []
    for task in TasksMib_Subgraph:
        benchmark = task.value.benchmark
        benchmark_cols = [
            display_mapping[f"{benchmark}_{model}"]  # Use display name from our mapping
            for model in task.value.models
            if f"{benchmark}_{model}" in dataframe.columns
        ]
        if benchmark_cols:
            benchmark_groups.append(benchmark_cols)
            print(f"\nBenchmark group for {benchmark}:", benchmark_cols)

    # Similarly for model groups
    model_groups = []
    all_models = list(set(model for task in TasksMib_Subgraph for model in task.value.models))
    
    for model in all_models:
        model_cols = [
            display_mapping[f"{task.value.benchmark}_{model}"]  # Use display name
            for task in TasksMib_Subgraph
            if model in task.value.models
            and f"{task.value.benchmark}_{model}" in dataframe.columns
        ]
        if model_cols:
            model_groups.append(model_cols)
            print(f"\nModel group for {model}:", model_cols)

    # Combine all groups using display names
    all_groups = benchmark_groups + model_groups
    all_columns = [col for group in all_groups for col in group]

    # Important: We need to rename our DataFrame columns to match display names
    
    renamed_df = dataframe.rename(columns=display_mapping)
    # all_columns = [c.name for c in fields(AutoEvalColumn_mib_subgraph) if c.displayed_by_default]
    # all_columns = [c.name for c in fields(AutoEvalColumn_mib_subgraph)]
    all_columns = renamed_df.columns.tolist()


    # Original code 
    return Leaderboard(
        value=renamed_df,  # Use DataFrame with display names
        datatype=[c.type for c in fields(AutoEvalColumn_mib_subgraph)],
        # select_columns=SelectColumns(
        #     default_selection=all_columns,  # Now contains display names
        #     label="Filter Results:",
        # ),
        search_columns=["Method"],
        hide_columns=["eval_name"],
        interactive=False,
    ), renamed_df
    



def init_leaderboard_mib_causalgraph(dataframe, track):
    # print("Debugging column issues:")
    # print("\nActual DataFrame columns:")
    # print(dataframe.columns.tolist())

    model_name_mapping = {
        "Qwen2ForCausalLM": "Qwen-2.5",
        "GPT2ForCausalLM": "GPT-2",
        "Gemma2ForCausalLM": "Gemma-2",
        "LlamaForCausalLM": "Llama-3.1"
    }

    benchmark_mapping = {
        "IOI": "IOI",
        "MCQA": "MCQA",
        "arithmetic_addition": "Arithmetic (+)",
        "arithmetic_subtraction": "Arithmetic (-)",
        "arc_easy": "ARC (Easy)",
        "arc_challenge": "ARC (Challenge)"
    }

    display_mapping = {}
    for task in TasksMib_Causalgraph:
        for model in task.value.models:
            field_name = f"{task.value.col_name}_{model}"
            display_name = f"{benchmark_mapping[task.value.col_name]} - {model_name_mapping[model]}"
            display_mapping[field_name] = display_name

    # print(dataframe)
    renamed_df = dataframe.rename(columns=display_mapping)
    
    # idx_to_method = {0: "Full Vector", 1: "DAS", 2: "DBM", 3: "PCA", 4: "SAE"}
    # idx_to_scores = {0: [0.38, 0.36, 0.38, 0.42],
    #                  1: [0.56, 0.62, 0.54, 0.51],
    #                  2: [0.43, 0.41, 0.53, 0.49],
    #                  3: [0.26, 0.20, 0.32, 0.40],
    #                  4: ["-", "-", 0.33, "-"]}
    # renamed_df.loc[0]["Method"] = "Full Vector"
    # for i in range(5):
    #     renamed_df.loc[i] = [idx_to_method[i]] + idx_to_scores[i]
        
    print(renamed_df)

    # Create only necessary columns
    return Leaderboard(
        value=renamed_df,
        datatype=[c.type for c in fields(AutoEvalColumn_mib_causalgraph)],
        # select_columns=SelectColumns(
        #     default_selection=["Method"],  # Start with just Method column
        #     cant_deselect=["Method"],      # Method column should always be visible
        #     label="Select Columns to Display:",
        # ),
        search_columns=["Method"],
        hide_columns=["eval_name"],
        bool_checkboxgroup_label="Hide models",
        interactive=False,
    ), renamed_df


def init_leaderboard(dataframe, track):
    if dataframe is None or dataframe.empty:
        raise ValueError("Leaderboard DataFrame is empty or None.")
    # filter for correct track
    dataframe = dataframe.loc[dataframe["Track"] == track]

    # print(f"\n\n\n dataframe is {dataframe}\n\n\n")
    
    return Leaderboard(
        value=dataframe,
        datatype=[c.type for c in fields(AutoEvalColumn)],
        select_columns=SelectColumns(
            default_selection=[c.name for c in fields(AutoEvalColumn) if c.displayed_by_default],
            cant_deselect=[c.name for c in fields(AutoEvalColumn) if c.never_hidden],
            label="Select Columns to Display:",
        ),
        search_columns=[AutoEvalColumn.model.name],
        hide_columns=[c.name for c in fields(AutoEvalColumn) if c.hidden],
        bool_checkboxgroup_label="Hide models",
        interactive=False,
    )

def process_json(temp_file):
    if temp_file is None:
        return {}

    # Handle file upload
    try:
        file_path = temp_file.name
        if file_path.endswith('.gz'):
            with gzip.open(file_path, 'rt') as f:
                data = json.load(f)
        else:
            with open(file_path, 'r') as f:
                data = json.load(f)
    except Exception as e:
        raise gr.Error(f"Error processing file: {str(e)}")

    gr.Markdown("Upload successful!")
    return data

def get_hf_username(hf_repo):
    hf_repo = hf_repo.rstrip("/")
    parts = hf_repo.split("/")
    username = parts[-2]
    return username


# Define the preset substrings for filtering
PRESET_SUBSTRINGS = ["IOI", "MCQA", "Arithmetic", "ARC", "GPT-2", "Qwen-2.5", "Gemma-2", "Llama-3.1"]
TASK_SUBSTRINGS = ["IOI", "MCQA", "Arithmetic", "ARC"]
MODEL_SUBSTRINGS = ["GPT-2", "Qwen-2.5", "Gemma-2", "Llama-3.1"]

def filter_columns_by_substrings(dataframe: pd.DataFrame, selected_task_substrings: List[str],
                                 selected_model_substrings: List[str]) -> pd.DataFrame:
    """
    Filter columns based on the selected substrings.
    """
    original_dataframe = deepcopy(dataframe)
    if not selected_task_substrings and not selected_model_substrings:
        return dataframe  # No filtering if no substrings are selected

    if not selected_task_substrings:
        # Filter columns that contain any of the selected model substrings
        filtered_columns = [
            col for col in dataframe.columns
            if any(sub.lower() in col.lower() for sub in selected_model_substrings)
            or col == "Method"
        ]
        return dataframe[filtered_columns]
    elif not selected_model_substrings:
        # Filter columns that contain any of the selected task substrings
        filtered_columns = [
            col for col in dataframe.columns
            if any(sub.lower() in col.lower() for sub in selected_task_substrings)
            or col == "Method"
        ]
        return dataframe[filtered_columns]

    # Filter columns by task first. Use AND logic to combine with model filtering
    filtered_columns = [
        col for col in dataframe.columns
        if any(sub.lower() in col.lower() for sub in selected_task_substrings)
        or col == "Method"
    ]
    filtered_columns = [
        col for col in dataframe[filtered_columns].columns
        if any(sub.lower() in col.lower() for sub in selected_model_substrings)
        or col == "Method"
    ]

    return dataframe[filtered_columns]

def update_leaderboard(dataframe: pd.DataFrame, selected_task_substrings: List[str],
                       selected_model_substrings: List[str]):
    """
    Update the leaderboard based on the selected substrings.
    """
    filtered_dataframe = filter_columns_by_substrings(dataframe, selected_task_substrings, selected_model_substrings)
    if len(selected_task_substrings) >= 2 or len(selected_task_substrings) == 0:
        if len(selected_model_substrings) >= 2 or len(selected_model_substrings) == 0:
            show_average = True
        else:
            show_average = False
    else:    
        show_average = False

    def _transform_floats(df):
        df_transformed = df.copy()
        # Apply transformation row by row
        for i, row in df_transformed.iterrows():
            # Apply sigmoid only to numeric values in the row
            df_transformed.loc[i] = row.apply(lambda x: _sigmoid(x) if isinstance(x, (float, int)) else x)
        return df_transformed

    if show_average:
        numeric_df = filtered_dataframe.select_dtypes(include=[np.number])
        means = numeric_df.mean(axis=1, skipna=False)
        # means = filtered_dataframe.replace("-", float("nan")).mean(axis=1, skipna=False)
        s_filtered_dataframe = _transform_floats(filtered_dataframe)
        numeric_s_df = s_filtered_dataframe.select_dtypes(include=[np.number])
        # s_means = s_filtered_dataframe.replace("-", float("nan")).mean(axis=1, skipna=False)
        s_means = numeric_s_df.mean(axis=1, skipna=False)
        filtered_dataframe.loc[:, "Average"] = np.where(filtered_dataframe.eq("-").any(axis=1), "-", means.round(2))
        filtered_dataframe.loc[:, "Score"] = np.where(filtered_dataframe.eq("-").any(axis=1), "-", s_means.round(2))
        filtered_dataframe = filtered_dataframe.sort_values(by=["Average"], ascending=False, na_position='last')

    # if show_average:
    #     print([row for index, row in filtered_dataframe.iterrows()])
    #     filtered_dataframe["Average"] = [round(np.mean(row.values()), 2) if "-" not in row.values() else "-" for index, row in filtered_dataframe.iterrows()]
    #     # Sort by Average score descending
    #     if 'Average' in dataframe.columns:
    #         # Convert '-' to NaN for sorting purposes
    #         df['Average'] = pd.to_numeric(['Average'], errors='coerce')
    #         df = df.sort_values(by=['Average'], ascending=True, na_position='last')
    #         # Convert NaN back to '-'
    #         df['Average'] = df['Average'].fillna('-')

    return filtered_dataframe

def process_url(url):
    # Add your URL processing logic here
    return f"You entered the URL: {url}"

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("Strict", elem_id="strict-benchmark-tab-table", id=0):
        #     leaderboard = init_leaderboard(LEADERBOARD_DF, "strict")
        # with gr.TabItem("Strict-small", elem_id="strict-small-benchmark-tab-table", id=1):
        #     leaderboard = init_leaderboard(LEADERBOARD_DF, "strict-small")
        # with gr.TabItem("Multimodal", elem_id="multimodal-benchmark-tab-table", id=2):
        #     leaderboard = init_leaderboard(LEADERBOARD_DF_MULTIMODAL, "multimodal")

        # with gr.TabItem("📝 About", elem_id="llm-benchmark-tab-table", id=4):
        #     gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
        
        # with gr.TabItem("Subgraph", elem_id="subgraph", id=0):
        #     leaderboard = init_leaderboard_mib_subgraph(LEADERBOARD_DF_MIB_SUBGRAPH, "Subgraph")
        with gr.TabItem("Circuit Localization", elem_id="subgraph", id=0):
            with gr.Tabs() as subgraph_tabs:
                with gr.TabItem("F+", id=0):
                    # Add description for filters
                    gr.Markdown("""
                    ### Filtering Options
                    Use the dropdown menus below to filter results by specific tasks or models. 
                    You can combine filters to see specific task-model combinations.
                    """)
                    # CheckboxGroup for selecting substrings
                    # substring_checkbox = gr.CheckboxGroup(
                    #     choices=PRESET_SUBSTRINGS,
                    #     label="Filter results:",
                    #     value=PRESET_SUBSTRINGS,  # Default to all substrings selected
                    # )
                    task_substring_checkbox = gr.CheckboxGroup(
                        choices=TASK_SUBSTRINGS,
                        label="View tasks:",
                        value=TASK_SUBSTRINGS,  # Default to all substrings selected
                    )
                    model_substring_checkbox = gr.CheckboxGroup(
                        choices = MODEL_SUBSTRINGS,
                        label = "View models:",
                        value = MODEL_SUBSTRINGS
                    )

                    leaderboard, data = init_leaderboard_mib_subgraph(LEADERBOARD_DF_MIB_SUBGRAPH_FPL, "Subgraph")
                    original_leaderboard = gr.State(value=data)
                    # Update the leaderboard when the user selects/deselects substrings
                    task_substring_checkbox.change(
                        fn=update_leaderboard,
                        inputs=[original_leaderboard, task_substring_checkbox, model_substring_checkbox],
                        outputs=leaderboard
                    )
                    model_substring_checkbox.change(
                        fn=update_leaderboard,
                        inputs=[original_leaderboard, task_substring_checkbox, model_substring_checkbox],
                        outputs=leaderboard
                    )
                    print(f"Leaderboard is {leaderboard}")
                with gr.TabItem("F=", id=1):
                    # Add description for filters
                    gr.Markdown("""
                    ### Filtering Options
                    Use the dropdown menus below to filter results by specific tasks or models. 
                    You can combine filters to see specific task-model combinations.
                    """)
                    # CheckboxGroup for selecting substrings
                    # substring_checkbox = gr.CheckboxGroup(
                    #     choices=PRESET_SUBSTRINGS,
                    #     label="Filter results:",
                    #     value=PRESET_SUBSTRINGS,  # Default to all substrings selected
                    # )
                    task_substring_checkbox = gr.CheckboxGroup(
                        choices=TASK_SUBSTRINGS,
                        label="View tasks:",
                        value=TASK_SUBSTRINGS,  # Default to all substrings selected
                    )
                    model_substring_checkbox = gr.CheckboxGroup(
                        choices = MODEL_SUBSTRINGS,
                        label = "View models:",
                        value = MODEL_SUBSTRINGS
                    )

                    leaderboard, data = init_leaderboard_mib_subgraph(LEADERBOARD_DF_MIB_SUBGRAPH_FEQ, "Subgraph")
                    original_leaderboard = gr.State(value=data)
                    # Update the leaderboard when the user selects/deselects substrings
                    task_substring_checkbox.change(
                        fn=update_leaderboard,
                        inputs=[original_leaderboard, task_substring_checkbox, model_substring_checkbox],
                        outputs=leaderboard
                    )
                    model_substring_checkbox.change(
                        fn=update_leaderboard,
                        inputs=[original_leaderboard, task_substring_checkbox, model_substring_checkbox],
                        outputs=leaderboard
                    )
                    print(f"Leaderboard is {leaderboard}")
    
        # Then modify the Causal Graph tab section
        with gr.TabItem("Causal Variable Localization", elem_id="causalgraph", id=1):
            with gr.Tabs() as causalgraph_tabs:
                with gr.TabItem("Detailed View", id=0):
                    leaderboard_detailed, data = init_leaderboard_mib_causalgraph(
                        LEADERBOARD_DF_MIB_CAUSALGRAPH_DETAILED, 
                        "Causal Graph"
                    )
                with gr.TabItem("Aggregated View", id=1):
                    gr.Markdown("""
                    ### Filtering Options
                    Use the dropdown menus below to filter results by specific tasks or models. 
                    You can combine filters to see specific task-model combinations.
                    """)
                    # substring_checkbox = gr.CheckboxGroup(
                    #     choices=PRESET_SUBSTRINGS,
                    #     label="Filter results:",
                    #     value=PRESET_SUBSTRINGS,  # Default to all substrings selected
                    # )
                    task_substring_checkbox = gr.CheckboxGroup(
                        choices=TASK_SUBSTRINGS,
                        label="View tasks:",
                        value=TASK_SUBSTRINGS,  # Default to all substrings selected
                    )
                    model_substring_checkbox = gr.CheckboxGroup(
                        choices = MODEL_SUBSTRINGS,
                        label = "View models:",
                        value = MODEL_SUBSTRINGS
                    )
                    
                    leaderboard_aggregated, data = init_leaderboard_mib_causalgraph(
                        LEADERBOARD_DF_MIB_CAUSALGRAPH_AGGREGATED, 
                        "Causal Graph"
                    )
                    original_leaderboard = gr.State(value=data)
                    task_substring_checkbox.change(
                        fn=update_leaderboard,
                        inputs=[original_leaderboard, task_substring_checkbox, model_substring_checkbox],
                        outputs=leaderboard_aggregated
                    )
                    model_substring_checkbox.change(
                        fn=update_leaderboard,
                        inputs=[original_leaderboard, task_substring_checkbox, model_substring_checkbox],
                        outputs=leaderboard_aggregated
                    )
                with gr.TabItem("Intervention Averaged", id=2):
                    leaderboard_averaged, data = init_leaderboard_mib_causalgraph(
                        LEADERBOARD_DF_MIB_CAUSALGRAPH_AVERAGED, 
                        "Causal Graph"
                    )
        
        with gr.TabItem("Submit", elem_id="llm-benchmark-tab-table", id=2):
            gr.Markdown("## 🏆 Submission Portal")
        
            # Track selection
            track = gr.Radio(
                choices=[
                    "Circuit Localization Track",
                    "Causal Variable Localization Track"
                ],
                label="Select Competition Track",
                elem_id="track_selector"
            )
            
            with gr.Group(visible=False) as circuit_ui:
                gr.Markdown("### Circuit Localization Requirements")
                hf_repo = gr.Textbox(
                    label="HuggingFace Repository URL",
                    placeholder="https://huggingface.co/username/repo/tree/main/path",
                    info="Must be a valid HuggingFace URL pointing to a folder with 10 circuit files (.json or .pt)"
                )
            
            with gr.Group(visible=False) as causal_ui:
                gr.Markdown("### Causal Variable Localization Requirements")
                with gr.Row():
                    layer = gr.Number(
                        label="Layer Number",
                        precision=0,
                        minimum=0,
                        info="Integer specifying the model layer"
                    )
                    token_position = gr.Number(
                        label="Token Position", 
                        precision=0,
                        minimum=0,
                        info="Integer specifying token position"
                    )
                code_upload = gr.File(
                    label="Upload Python file implementing your featurization function",
                    file_types=[".py"],
                )
            
            # Common fields
            with gr.Group():
                gr.Markdown("### Team Information")
                team_name = gr.Textbox(label="Team Name")
                contact_email = gr.Textbox(label="Contact Email")

            # Dynamic UI logic
            def toggle_ui(track):
                circuit = track == "Circuit Localization Track"
                causal = not circuit
                return {
                    circuit_ui: gr.Group(visible=circuit),
                    causal_ui: gr.Group(visible=causal)
                }
            
            track.change(toggle_ui, track, [circuit_ui, causal_ui])
            
            # Submission handling
            status = gr.Textbox(label="Submission Status", visible=False)
            
            def handle_submission(track, hf_repo, layer, token_position, code_upload, team_name, contact_email):
                errors = []
                
                # Validate common fields
                if not team_name.strip():
                    errors.append("Team name is required")
                if "@" not in contact_email or "." not in contact_email:
                    errors.append("Valid email address is required")
                
                # Track-specific validation
                if "Circuit" in track:
                    if not hf_repo.startswith("https://huggingface.co/"):
                        errors.append("Invalid HuggingFace URL - must start with https://huggingface.co/")
                    else:
                        # Check rate limit only for valid HF submissions
                        username = get_hf_username(hf_repo)
                        rate = 0  # TODO: check submissions queue for rates
                        rate_limit = 2
                        if rate > rate_limit:
                            errors.append("Rate limit exceeded (max 2 submissions per week per HF account)")
            
                else:
                    if not (isinstance(layer, int) and isinstance(token_position, int)):
                        errors.append("Layer and token position must be integers")
                    if not code_upload:
                        errors.append("Code file upload is required")
                
                if errors:
                    return gr.Textbox("\n".join(f"❌ {e}" for e in errors), visible=True)
                
                # Process valid submission
                return gr.Textbox("✅ Submission received! Thank you for your entry.", visible=True)
            
            submit_btn = gr.Button("Submit Entry", variant="primary")
            submit_btn.click(
                handle_submission,
                inputs=[track, hf_repo, layer, token_position, code_upload, team_name, contact_email],
                outputs=status
            )
            
            # Add info about rate limits
            gr.Markdown("""
            ### Submission Policy
            - Maximum 2 valid submissions per HuggingFace account per week
            - Invalid submissions don't count toward your limit
            - Rate limit tracked on a rolling basis: a submission no longer counts toward the limit as soon as 7 days have passed since the submission time
            """)

    # with gr.Row():
    #     with gr.Accordion("📙 Citation", open=False):
    #         citation_button = gr.Textbox(
    #             value=CITATION_BUTTON_TEXT,
    #             label=CITATION_BUTTON_LABEL,
    #             lines=20,
    #             elem_id="citation-button",
    #             show_copy_button=True,
    #         )

scheduler = BackgroundScheduler()
scheduler.add_job(restart_space, "interval", seconds=1800)
scheduler.start()
demo.launch(share=True, ssr_mode=False)