src/populate.py

import json
import os

import pandas as pd

from src.display.formatting import has_no_nan_values, make_clickable_model
from src.display.utils import AutoEvalColumn, EvalQueueColumn
from src.leaderboard.read_evals import get_raw_eval_results
from src.about import Tasks


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

    df = pd.DataFrame.from_records(all_data_json)
    # Sort by accuracy first, then F1 score
    if len(df):
        df = df.sort_values(by=[Tasks.task0.value.col_name, Tasks.task1.value.col_name], ascending=False)
        df[benchmark_cols] = df[benchmark_cols].round(decimals=3)
    
        # Filter out incomplete submissions
        df = df[has_no_nan_values(df, benchmark_cols)]

    return df