src/populate.py

import json
import os

import pandas as pd
from datasets import load_dataset, get_dataset_config_names
from datasets.exceptions import DatasetNotFoundError
from tqdm.auto import tqdm

from src.display.formatting import has_no_nan_values, make_clickable_model
from src.display.utils import AutoEvalColumn, EvalQueueColumn
from src.envs import TOKEN
from src.leaderboard.read_evals import get_raw_eval_results
from src.logger import get_logger

logger = get_logger(__name__)


def get_leaderboard_df(results_dataset_name: str) -> pd.DataFrame:
    """Creates a dataframe from all the individual experiment results"""

    try:
        configs = get_dataset_config_names(results_dataset_name, token=TOKEN)
    except (DatasetNotFoundError, FileNotFoundError):
        # Return an empty DataFrame with expected columns
        return pd.DataFrame(
            columns=[
                "System Name",
                "System Type",
                "Organization",
                "Success Rate (%)",
                "Problems Solved",
                "Submitted On",
            ]
        )

    rows = []
    for submission_id in tqdm(configs, total=len(configs), desc="Processing Submission Results"):
        submission_ds = load_dataset(results_dataset_name, submission_id, split="train", token=TOKEN)
        submission_df = pd.DataFrame(submission_ds)

        if submission_df.empty or "did_pass" not in submission_df.columns or submission_df.did_pass.isna().any():
            logger.warning(f"Skipping {submission_id} due to invalid did_pass values")
            continue

        success_rate = 100 * submission_df["did_pass"].mean()
        num_solved = submission_df["did_pass"].sum()
        first_row = submission_df.iloc[0]

        rows.append(
            {
                "System Name": first_row["system_name"],
                "System Type": first_row["system_type"],
                "Organization": first_row["organization"],
                "Success Rate (%)": success_rate,
                "Problems Solved": num_solved,
                "Submitted On": pd.to_datetime(first_row["submission_ts"]).strftime("%Y-%m-%d %H:%M"),
            }
        )

    full_df = pd.DataFrame(rows)

    # TODO: forbid multiple submissions under the same name?
    # Keep only the latest entry per unique (System Name, System Type, Organization) triplet
    final_df = (
        full_df.sort_values("Submitted On", ascending=False)
        .drop_duplicates(subset=["System Name", "System Type", "Organization"], keep="first")
        .sort_values(by=[AutoEvalColumn.success_rate.name], ascending=False)
        .reset_index(drop=True)
    )

    cols_to_round = ["Success Rate (%)"]
    final_df[cols_to_round] = final_df[cols_to_round].round(decimals=2)

    return final_df


def get_evaluation_queue_df(save_path: str, cols: list) -> list[pd.DataFrame]:
    """Creates the different dataframes for the evaluation queues requestes"""
    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 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", "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]