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.filter_models import filter_models
from src.leaderboard.read_evals import get_raw_eval_results, EvalResult


def get_leaderboard_df(results_path: str, requests_path: str, cols: list, benchmark_cols: list) -> tuple[list[EvalResult], pd.DataFrame]:
    # Returns a list of EvalResult
    # raw_data[0]:
    # EvalResult(eval_name='EleutherAI_pythia-1.3b_torch.float16', full_model='EleutherAI/pythia-1.3b', org='EleutherAI', model='pythia-1.3b', revision='34b668ff0acfe56f2d541aa46b385557ee39eb3f', results={'arc:challenge': 31.14334470989761, 'hellaswag': 51.43397729535949, 'hendrycksTest': 26.55151159544371, 'truthfulqa:mc': 39.24322830092449, 'winogrande': 57.37963693764798, 'gsm8k': 0.9855951478392722, 'drop': 4.056312919463095}, precision='torch.float16', model_type=<ModelType.PT: ModelTypeDetails(name='pretrained', symbol='🟢')>, weight_type='Original', architecture='GPTNeoXForCausalLM', license='apache-2.0', likes=7, num_params=1.312, date='2023-09-09T10:52:17Z', still_on_hub=True)
    # EvalResult and get_raw_eval_results are defined in ./src/leaderboard/read_evals.py, the results slots are not hardcoded
    raw_data = get_raw_eval_results(results_path, requests_path)
    all_data_json = [v.to_dict() for v in raw_data]
    # all_data_json.append(baseline_row)
    filter_models(all_data_json)

    df = pd.DataFrame.from_records(all_data_json)
    if AutoEvalColumn.average.name in df:
        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)]
    return raw_data, df


def get_evaluation_queue_df(save_path: str, cols: list) -> tuple[pd.DataFrame, pd.DataFrame, 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]