rodrigomasini/data_only_hallucination_leaderboard

1import json2import os3 4import pandas as pd5 6from src.display.formatting import has_no_nan_values, make_clickable_model7from src.display.utils import AutoEvalColumn, EvalQueueColumn8from src.leaderboard.filter_models import filter_models9from src.leaderboard.read_evals import get_raw_eval_results, EvalResult10 11 12def get_leaderboard_df(results_path: str, requests_path: str, cols: list, benchmark_cols: list) -> tuple[list[EvalResult], pd.DataFrame]:13    # Returns a list of EvalResult14    # raw_data[0]:15    # 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)16    # EvalResult and get_raw_eval_results are defined in ./src/leaderboard/read_evals.py, the results slots are not hardcoded17    raw_data = get_raw_eval_results(results_path, requests_path)18    all_data_json = [v.to_dict() for v in raw_data if v.is_complete()]19    # all_data_json.append(baseline_row)20    filter_models(all_data_json)21 22    df = pd.DataFrame.from_records(all_data_json)23    if AutoEvalColumn.average.name in df:24        df = df.sort_values(by=[AutoEvalColumn.average.name], ascending=False)25        df = df[cols].round(decimals=2)26 27        # filter out if any of the benchmarks have not been produced28        df = df[has_no_nan_values(df, benchmark_cols)]29    return raw_data, df30 31 32def get_evaluation_queue_df(save_path: str, cols: list) -> tuple[pd.DataFrame, pd.DataFrame, pd.DataFrame]:33    entries = [entry for entry in os.listdir(save_path) if not entry.startswith(".")]34    all_evals = []35 36    for entry in entries:37        if ".json" in entry:38            file_path = os.path.join(save_path, entry)39            with open(file_path) as fp:40                data = json.load(fp)41 42            data[EvalQueueColumn.model.name] = make_clickable_model(data["model"])43            data[EvalQueueColumn.revision.name] = data.get("revision", "main")44 45            all_evals.append(data)46        elif ".md" not in entry:47            # this is a folder48            sub_entries = [e for e in os.listdir(f"{save_path}/{entry}") if not e.startswith(".")]49            for sub_entry in sub_entries:50                file_path = os.path.join(save_path, entry, sub_entry)51                with open(file_path) as fp:52                    data = json.load(fp)53 54                data[EvalQueueColumn.model.name] = make_clickable_model(data["model"])55                data[EvalQueueColumn.revision.name] = data.get("revision", "main")56                all_evals.append(data)57 58    pending_list = [e for e in all_evals if e["status"] in ["PENDING", "RERUN"]]59    running_list = [e for e in all_evals if e["status"] == "RUNNING"]60    finished_list = [e for e in all_evals if e["status"].startswith("FINISHED") or e["status"] == "PENDING_NEW_EVAL"]61    df_pending = pd.DataFrame.from_records(pending_list, columns=cols)62    df_running = pd.DataFrame.from_records(running_list, columns=cols)63    df_finished = pd.DataFrame.from_records(finished_list, columns=cols)64    return df_finished[cols], df_running[cols], df_pending[cols]65