import json import os import copy import pandas as pd from src.display.formatting import has_no_nan_values, make_requests_clickable_model from src.display.utils import AutoEvalColumn, EvalQueueColumn, baseline_row, external_rows from src.leaderboard.filter_models import filter_models_flags from src.leaderboard.read_evals import get_raw_eval_results def get_leaderboard_df(results_path: str, requests_path: str, dynamic_path: str, cols: list, benchmark_cols: list, show_incomplete=False) -> pd.DataFrame: raw_data = get_raw_eval_results(results_path=results_path, requests_path=requests_path, dynamic_path=dynamic_path) all_data_json = [v.to_dict() for v in raw_data] all_data_json.append(baseline_row) for external_row in external_rows: all_data_json.append(external_row) filter_models_flags(all_data_json) 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 if not show_incomplete: df = df[has_no_nan_values(df, benchmark_cols)] return raw_data, df def get_evaluation_queue_df(save_path: str, cols: list, show_incomplete=False) -> list[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_requests_clickable_model(data["model"], entry) 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_requests_clickable_model(data["model"], os.path.join(entry, sub_entry)) data[EvalQueueColumn.revision.name] = data.get("revision", "main") all_evals.append(data) cols_pending = copy.deepcopy(cols) cols_pending.append('source') cols_pending.append('submitted_time') pending_list = [e for e in all_evals if e["status"] in ["PENDING", "RERUN", "PENDING_NEW_EVAL"]] running_list = [e for e in all_evals if e["status"] == "RUNNING"] finished_list = [e for e in all_evals if e["status"] in ["FINISHED", "PENDING_NEW_EVAL" if show_incomplete else "FINISHED"]] failed_list = [e for e in all_evals if e["status"] == "FAILED"] df_pending = pd.DataFrame.from_records(pending_list, columns=cols_pending) df_running = pd.DataFrame.from_records(running_list, columns=cols) df_finished = pd.DataFrame.from_records(finished_list, columns=cols) df_failed = pd.DataFrame.from_records(failed_list, columns=cols) df_pending['source_priority'] = df_pending["source"].apply(lambda x: {"manual": 0, "leaderboard": 1, "script": 2}.get(x, 3)) df_pending['status_priority'] = df_pending["status"].apply(lambda x: {"PENDING": 2, "RERUN": 0, "PENDING_NEW_EVAL": 1}.get(x, 3)) df_pending = df_pending.sort_values(['source_priority', 'status_priority', 'submitted_time']) df_pending = df_pending.drop(['source_priority', 'status_priority', 'submitted_time', 'source'], axis=1) return df_finished[cols], df_running[cols], df_pending[cols], df_failed[cols]