src/populate.py

import json
import os

import numpy as np
import pandas as pd

from src.about import Tasks
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


TASK_NAME_INVERSE_MAP = dict()
for task in Tasks:
    TASK_NAME_INVERSE_MAP[task.value.col_name] = {
        "name": task.value.benchmark,
        "type": task.value.type,
        "source": task.value.source,
    }

EMPTY_SYMBOL = "&ndash"


def get_inspect_log_url(model_name: str, benchmark_name: str) -> str:
    """Returns the URL to the log file for a given model and benchmark"""
    with open("./inspect_log_file_names.json", "r") as f:
        inspect_log_files = json.load(f)
    log_file_name = inspect_log_files[model_name].get(benchmark_name, None)
    if log_file_name is None:
        return ""
    else:
        # replace .json with .eval
        log_file_name = log_file_name.replace(".json", ".eval")
        return f"https://storage.googleapis.com/inspect-evals/{model_name}/index.html?log_file=logs/logs/{log_file_name}"


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)

    # df = df.sort_values(by=[AutoEvalColumn.average.name], ascending=False)
    df = df[cols].round(decimals=2)

    # # filter out if any of the benchmarks have not been produced
    # df = df[has_no_nan_values(df, benchmark_cols)]
    df = df.fillna(EMPTY_SYMBOL)
    print(df["GAIA"].head())

    # make values clickable and link to log files
    for col in benchmark_cols:
        df[col] = df[[AutoEvalColumn.model.name, col]].apply(lambda x: f"[{x[col]}]({get_inspect_log_url(model_name=x[AutoEvalColumn.model.name].split('>')[1].split('<')[0], benchmark_name=TASK_NAME_INVERSE_MAP[col]['name'])})" if x[col] != EMPTY_SYMBOL else x[col], axis=1)

    # # make task names clickable and link to inspect-evals repository - this creates issues later
    # df = df.rename(columns={col: f"[{col}]({TASK_NAME_INVERSE_MAP[col]['source']})" for col in benchmark_cols})

    return 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]