src/populate.py

import json
import os
import numpy as np
import pandas as pd

from src.display.formatting import has_no_nan_values, make_clickable_model, model_hyperlink
from src.display.formatting import has_no_nan_values, make_clickable_model
from src.display.utils import AutoEvalColumn, EvalQueueColumn, HEADER_MAP
from src.leaderboard.read_evals import get_raw_eval_results


def get_leaderboard_df(results_path: str, requests_path: str, cols: list, benchmark_cols: list) -> pd.DataFrame:
    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(raw_data)
    #df = df.sort_values(by=[AutoEvalColumn.average.name], ascending=False)
    df = df.rename(columns=HEADER_MAP)
    df[df.select_dtypes(include=['number']).columns] *= 100 # convert to percentage
    df["Grammar (Avg.)"] = df[["AGREE"]].mean(axis=1)
    df["Knowledge (Avg.)"] = df[["ARC-Challenge", "ARC-Easy", "MMLU", "TruthfulQA"]].mean(axis=1)
    df["Reasoning (Avg.)"] = df[["ANLI", "Belebele", "CTKFacts", "SQAD"]].mean(axis=1)
    df["Math (Avg.)"] = df[["GSM8K", "Klokanek"]].mean(axis=1)
    df["Classification (Avg.)"] = df[["Czech News", "Facebook Comments", "Mall Reviews", "Subjectivity"]].mean(axis=1)
    df["Aggregate Score"] = df[["Grammar (Avg.)", "Knowledge (Avg.)", "Reasoning (Avg.)", "Math (Avg.)", "Classification (Avg.)"]].mean(axis=1)
    df[" "] = ""  # The dataframe does not display the last column - BUG in gradio?
    df = df[cols].round(decimals=2)
    df.replace(r'\s+', np.nan, regex=True)
    # filter out if any of the benchmarks have not been produced
    df = df[has_no_nan_values(df, benchmark_cols)]
    
    return raw_data, df, 


def get_evaluation_queue_df(save_path: str, cols: list) -> list[pd.DataFrame]:
    entries = [entry for entry in os.listdir(save_path) if not entry.startswith(".")]
    all_evals = []
    
    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)

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

            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 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", "RERUN"]]
    running_list = [e for e in all_evals if e["status"] == "RUNNING"]
    finished_list = [e for e in all_evals if e["status"].startswith("FINISHED") or e["status"] == "PENDING_NEW_EVAL"]
    df_pending = pd.DataFrame.from_records(pending_list, columns=cols)
    df_running = pd.DataFrame.from_records(running_list, columns=cols)
    df_finished = pd.DataFrame.from_records(finished_list, columns=cols)
    return df_finished[cols], df_running[cols], df_pending[cols]