src/populate.py

import json
import os
import pandas as pd
from typing import List, Dict, Tuple

from src.display.formatting import has_no_nan_values, make_clickable_model
from src.display.utils import AutoEvalColumn, AutoEvalColumnMultimodal, EvalQueueColumn
from src.leaderboard.read_evals import get_raw_eval_results, get_raw_eval_results_mib_subgraph, get_raw_eval_results_mib_causalgraph
from src.about import TasksMib_Causalgraph
from src.submission.check_validity import parse_huggingface_url

def get_leaderboard_df_mib_subgraph(results_path: str, cols: list, benchmark_cols: list,
                                    metric_type = "CPR") -> pd.DataFrame:
    """Creates a dataframe from all the MIB experiment results"""
    # print(f"results_path is {results_path}, requests_path is {requests_path}")
    raw_data = get_raw_eval_results_mib_subgraph(results_path)

    all_data_json = [v.to_dict(metric_type=metric_type) for v in raw_data]
    # print(f"all_data_json is {pd.DataFrame.from_records(all_data_json)}")

    # Convert to dataframe
    df = pd.DataFrame.from_records(all_data_json)
    ascending = False if metric_type == "CPR" else True
    
    # Sort by Average score descending
    if 'Average' in df.columns:
        # Convert '-' to NaN for sorting purposes
        df['Average'] = pd.to_numeric(df['Average'], errors='coerce')
        df = df.sort_values(by=['Average'], ascending=ascending, na_position='last')
        # Convert NaN back to '-'
        df['Average'] = df['Average'].fillna('-')
    
    return df



def aggregate_methods(df: pd.DataFrame) -> pd.DataFrame:
    """Aggregates rows with the same base method name by taking the max value for each column"""
    df_copy = df.copy()
    
    # Set Method as index if it isn't already
    if 'Method' in df_copy.columns:
        df_copy.set_index('Method', inplace=True)
    
    # Extract base method names (remove _2, _3, etc. suffixes)
    base_methods = [name.split('_')[0] if '_' in str(name) and str(name).split('_')[-1].isdigit() 
                   else name for name in df_copy.index]
    df_copy.index = base_methods
    
    # Convert scores to numeric values
    numeric_df = df_copy.select_dtypes(include=['float64', 'int64'])
    
    # Group by base method name and take the max
    aggregated_df = numeric_df.groupby(level=0).max().round(2)
    
    # Reset index to get Method as a column
    aggregated_df.reset_index(inplace=True)
    aggregated_df.rename(columns={'index': 'Method'}, inplace=True)
    
    return aggregated_df


def create_intervention_averaged_df(df: pd.DataFrame) -> pd.DataFrame:
    """Creates a DataFrame where columns are model_task and cells are averaged over interventions"""
    df_copy = df.copy()
    
    # Store Method column
    method_col = None
    if 'Method' in df_copy.columns:
        method_col = df_copy['Method']
        df_copy = df_copy.drop('Method', axis=1)
    
    if 'eval_name' in df_copy.columns:
        df_copy = df_copy.drop('eval_name', axis=1)
    
    # Group columns by model and intervention
    result_cols = {}
    for task in TasksMib_Causalgraph:
        for model in task.value.models:  # Will iterate over all three models
            for intervention in task.value.interventions:
                for counterfactual in task.value.counterfactuals:
                    col_pattern = f"{model}_layer.*_{intervention}_{counterfactual}"
                    matching_cols = [c for c in df_copy.columns if pd.Series(c).str.match(col_pattern).any()]
                    if matching_cols:
                        col_name = f"{model}_{intervention}_{counterfactual}"
                        result_cols[col_name] = matching_cols
    
    averaged_df = pd.DataFrame()
    if method_col is not None:
        averaged_df['Method'] = method_col
    
    for col_name, cols in result_cols.items():
        averaged_df[col_name] = df_copy[cols].mean(axis=1).round(2)
    
    return averaged_df



def get_leaderboard_df_mib_causalgraph(results_path: str) -> Tuple[pd.DataFrame, pd.DataFrame, pd.DataFrame]:


    aggregated_df, intervention_averaged_df = get_raw_eval_results_mib_causalgraph(results_path)

    print(f"Columns in aggregated_df: {aggregated_df.columns.tolist()}")
    print(f"Columns in intervention_averaged_df: {intervention_averaged_df.columns.tolist()}")
    
    return aggregated_df, intervention_averaged_df


def get_evaluation_queue_df(save_path: str, cols: list, track: str) -> list[pd.DataFrame]:
    """Creates the different dataframes for the evaluation queues requests"""
    entries = [entry for entry in os.listdir(save_path) if not entry.startswith(".")]
    all_evals = []

    print(track)

    for entry in entries:
        if ".json" in entry:
            file_path = os.path.join(save_path, entry)
            with open(file_path) as fp:
                data = json.load(fp)

            # if "still_on_hub" in data and data["still_on_hub"]:
            #     data[EvalQueueColumn.model.name] = make_clickable_model(data["hf_repo"], data["model"])
            #     data[EvalQueueColumn.revision.name] = data.get("revision", "main")
            # else:
            #     data[EvalQueueColumn.model.name] = data["model"]
            #     data[EvalQueueColumn.revision.name] = "N/A"

            all_evals.append(data)

        # elif ".md" not in entry:
        #     # this is a folder
        #     sub_entries = [e for e in os.listdir(f"{save_path}/{entry}") if os.path.isfile(e) and not e.startswith(".")]
        #     for sub_entry in sub_entries:
        #         file_path = os.path.join(save_path, entry, sub_entry)
        #         with open(file_path) as fp:
        #             data = json.load(fp)

        #         data[EvalQueueColumn.model.name] = make_clickable_model(data["model"])
        #         data[EvalQueueColumn.revision.name] = data.get("revision", "main")
        #         all_evals.append(data)

    pending_list = [e for e in all_evals if e["status"] in ["PENDING", "PREVALIDATION"]]
    finished_list = [e for e in all_evals if e["status"].startswith("FINISHED") or e["status"] == "PENDING_NEW_EVAL" or e["status"] == "FAILED"]
    for list in (pending_list, finished_list):
        for item in list:
            item["track"] = track
            item["hf_repo"] = parse_huggingface_url(item["hf_repo"])[0]
    df_pending = pd.DataFrame.from_records(pending_list, columns=cols)
    df_finished = pd.DataFrame.from_records(finished_list, columns=cols)
    return df_finished[cols], df_pending[cols]