src/populate.py

import pandas as pd
from src.display.formatting import has_no_nan_values, make_clickable_model
from src.display.utils import AutoEvalColumn
from src.leaderboard.read_evals import get_raw_eval_results

def get_leaderboard_df(results_path: str, cols: list, benchmark_cols: list) -> pd.DataFrame:
    """Creates a dataframe from all the individual experiment results"""
    try:
        raw_data = get_raw_eval_results(results_path)
        all_data_json = [v.to_dict() for v in raw_data]

        if not all_data_json:
            # Create empty DataFrame with correct columns
            empty_df = pd.DataFrame(columns=cols)
            # Ensure correct column types
            empty_df[AutoEvalColumn.average.name] = pd.Series(dtype=float)
            for col in benchmark_cols:
                empty_df[col] = pd.Series(dtype=float)
            return empty_df

        df = pd.DataFrame.from_records(all_data_json)
        df = df.sort_values(by=[AutoEvalColumn.average.name], ascending=False)
        df = df[cols].round(decimals=2)

        # filter out if perplexity hasn't been evaluated
        df = df[has_no_nan_values(df, benchmark_cols)]
        return df
    except Exception as e:
        print(f"Error creating leaderboard: {e}")
        # Return empty DataFrame with correct structure
        empty_df = pd.DataFrame(columns=cols)
        empty_df[AutoEvalColumn.average.name] = pd.Series(dtype=float)
        for col in benchmark_cols:
            empty_df[col] = pd.Series(dtype=float)
        return empty_df