open_llm_leaderboard / src /populate.py
Alina Lozovskaia
Updated collections.py
2293858
raw
history blame
2.87 kB
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, baseline_row
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
) -> 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)
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
df = df[has_no_nan_values(df, benchmark_cols)]
return raw_data, df
def get_evaluation_queue_df(save_path: str, cols: list) -> 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_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:
try:
data = json.load(fp)
except json.JSONDecodeError:
print(f"Error reading {file_path}")
continue
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]