MCILAB/LLM_Alignment_Evaluation

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.read_evals import get_raw_eval_results9 10 11def get_leaderboard_df(results_path: str, requests_path: str, cols: list, benchmark_cols: list) -> pd.DataFrame:12    """Creates a dataframe from all the individual experiment results"""13    raw_data = get_raw_eval_results(results_path, requests_path)14    all_data_json = [v.to_dict() for v in raw_data]15    df = pd.DataFrame.from_records(all_data_json)16    df = df[cols].round(decimals=2)17    # filter out if any of the benchmarks have not been produced18    df = df[has_no_nan_values(df, benchmark_cols)]19    return df20 21 22def get_evaluation_queue_df(save_path: str, cols: list) -> list[pd.DataFrame]:23    """Creates the different dataframes for the evaluation queues requestes"""24    entries = [entry for entry in os.listdir(save_path) if not entry.startswith(".")]25    all_evals = []26 27    for entry in entries:28        if ".json" in entry:29            file_path = os.path.join(save_path, entry)30            with open(file_path) as fp:31                data = json.load(fp)32 33            data[EvalQueueColumn.model.name] = make_clickable_model(data["model"])34            data[EvalQueueColumn.revision.name] = data.get("revision", "main")35 36            all_evals.append(data)37        elif ".md" not in entry:38            # this is a folder39            sub_entries = [e for e in os.listdir(f"{save_path}/{entry}") if os.path.isfile(e) and not e.startswith(".")]40            for sub_entry in sub_entries:41                file_path = os.path.join(save_path, entry, sub_entry)42                with open(file_path) as fp:43                    data = json.load(fp)44 45                data[EvalQueueColumn.model.name] = make_clickable_model(data["model"])46                data[EvalQueueColumn.revision.name] = data.get("revision", "main")47                all_evals.append(data)48 49    pending_list = [e for e in all_evals if e["status"] in ["PENDING", "RERUN"]]50    running_list = [e for e in all_evals if e["status"] == "RUNNING"]51    finished_list = [e for e in all_evals if e["status"].startswith("FINISHED") or e["status"] == "PENDING_NEW_EVAL"]52    df_pending = pd.DataFrame.from_records(pending_list, columns=cols)53    df_running = pd.DataFrame.from_records(running_list, columns=cols)54    df_finished = pd.DataFrame.from_records(finished_list, columns=cols)55    return df_finished[cols], df_running[cols], df_pending[cols]56