eval-leaderboard / create_log_file_map.py
xeon27
Add script for log file map creation
ca5fb3d
raw
history blame
1.43 kB
import json
import os
from collections import defaultdict
def main():
base_bm_input_path = "./base_benchmarking_logs"
agentic_bm_input_path = "./agentic_benchmarking_logs"
log_file_map = defaultdict()
for model_name in os.listdir(base_bm_input_path):
log_file_map[model_name] = defaultdict(str)
if os.path.isdir(os.path.join(base_bm_input_path, model_name)):
for task_log_file in os.listdir(os.path.join(base_bm_input_path, model_name)):
with open(os.path.join(base_bm_input_path, model_name, task_log_file), "r") as f:
result = json.load(f)
task_name = result["eval"]["task"].split("/")[-1]
log_file_map[model_name][task_name] = task_log_file
for model_name in os.listdir(agentic_bm_input_path):
if os.path.isdir(os.path.join(agentic_bm_input_path, model_name)):
for task_log_file in os.listdir(os.path.join(agentic_bm_input_path, model_name)):
with open(os.path.join(agentic_bm_input_path, model_name, task_log_file), "r") as f:
result = json.load(f)
task_name = result["eval"]["task"].split("/")[-1]
log_file_map[model_name][task_name] = task_log_file
with open("./inspect_log_file_names.json", "w") as f:
json.dump(log_file_map, f, indent=4)
if __name__ == "__main__":
main()