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.read_evals import get_raw_eval_results 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) # ------------------------------------------------------------------ # Fallback: if no evaluation results are found we populate the # leaderboard with a single example model. This guarantees that a # freshly deployed Space shows a non-empty leaderboard and it serves # as a template for the expected columns/values. # ------------------------------------------------------------------ if df.empty: example_row = {} # Populate benchmark metrics with the default value 0.5 using internal column names for metric in benchmark_cols: example_row[metric] = 0.5 # Minimal metadata so that the row displays nicely example_row[AutoEvalColumn.model.name] = make_clickable_model("example/model") example_row[AutoEvalColumn.average.name] = 0.5 example_row[AutoEvalColumn.model_type_symbol.name] = "🟢" example_row[AutoEvalColumn.model_type.name] = "pretrained" example_row[AutoEvalColumn.precision.name] = "float16" example_row[AutoEvalColumn.weight_type.name] = "Original" example_row[AutoEvalColumn.still_on_hub.name] = True example_row[AutoEvalColumn.architecture.name] = "Transformer" example_row[AutoEvalColumn.revision.name] = "main" example_row[AutoEvalColumn.license.name] = "apache-2.0" # Any missing columns will be created later in the function df = pd.DataFrame([example_row]) # Sort primarily by LLM exact-match Pass@1 metric; if not present, fall back to average preferred_cols = [] if hasattr(AutoEvalColumn, "pass_at_1"): preferred_cols.append(AutoEvalColumn.pass_at_1.name) preferred_cols.append(AutoEvalColumn.average.name) for col in preferred_cols: if col in df.columns: df = df.sort_values(by=[col], ascending=False) break # Ensure all expected columns exist, add missing ones with NaN so selection does not fail for expected in cols: if expected not in df.columns: df[expected] = pd.NA 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 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]