leaderboard / backend-cli.py
pminervini's picture
update
6c79b12
raw
history blame
3.56 kB
import os
import json
from datetime import datetime
from huggingface_hub import snapshot_download
from src.backend.run_eval_suite import run_evaluation
from src.backend.manage_requests import check_completed_evals, get_eval_requests, set_eval_request
from src.backend.sort_queue import sort_models_by_priority
from src.backend.envs import Tasks, EVAL_REQUESTS_PATH_BACKEND,EVAL_RESULTS_PATH_BACKEND, DEVICE, LIMIT
from src.envs import QUEUE_REPO, RESULTS_REPO, API
import logging
import pprint
# TASKS_HARNESS = [task.value.benchmark for task in Tasks]
logging.getLogger("openai").setLevel(logging.WARNING)
logging.basicConfig(level=logging.ERROR)
pp = pprint.PrettyPrinter(width=80)
PENDING_STATUS = "PENDING"
RUNNING_STATUS = "RUNNING"
FINISHED_STATUS = "FINISHED"
FAILED_STATUS = "FAILED"
snapshot_download(repo_id=RESULTS_REPO, revision="main", local_dir=EVAL_RESULTS_PATH_BACKEND, repo_type="dataset", max_workers=60)
snapshot_download(repo_id=QUEUE_REPO, revision="main", local_dir=EVAL_REQUESTS_PATH_BACKEND, repo_type="dataset", max_workers=60)
def run_auto_eval():
current_pending_status = [PENDING_STATUS]
# pull the eval dataset from the hub and parse any eval requests
# check completed evals and set them to finished
check_completed_evals(api=API, checked_status=RUNNING_STATUS, completed_status=FINISHED_STATUS,
failed_status=FAILED_STATUS, hf_repo=QUEUE_REPO, local_dir=EVAL_REQUESTS_PATH_BACKEND,
hf_repo_results=RESULTS_REPO, local_dir_results=EVAL_RESULTS_PATH_BACKEND)
# Get all eval request that are PENDING, if you want to run other evals, change this parameter
eval_requests = get_eval_requests(job_status=current_pending_status, hf_repo=QUEUE_REPO, local_dir=EVAL_REQUESTS_PATH_BACKEND)
# Sort the evals by priority (first submitted first run)
eval_requests = sort_models_by_priority(api=API, models=eval_requests)
print(f"Found {len(eval_requests)} {','.join(current_pending_status)} eval requests")
if len(eval_requests) == 0:
return
eval_request = eval_requests[0]
pp.pprint(eval_request)
set_eval_request(api=API, eval_request=eval_request, set_to_status=RUNNING_STATUS, hf_repo=QUEUE_REPO,
local_dir=EVAL_REQUESTS_PATH_BACKEND)
# results = run_evaluation(eval_request=eval_request, task_names=TASKS_HARNESS, num_fewshot=NUM_FEWSHOT,
# batch_size=1, device=DEVICE, no_cache=True, limit=LIMIT)
TASKS_HARNESS = [task.value for task in Tasks]
for task in TASKS_HARNESS:
results = run_evaluation(eval_request=eval_request, task_names=[task.benchmark], num_fewshot=task.num_fewshot,
batch_size=1, device=DEVICE, no_cache=True, limit=LIMIT)
dumped = json.dumps(results, indent=2)
print(dumped)
output_path = os.path.join(EVAL_RESULTS_PATH_BACKEND, *eval_request.model.split("/"), f"results_{datetime.now()}.json")
os.makedirs(os.path.dirname(output_path), exist_ok=True)
with open(output_path, "w") as f:
f.write(dumped)
API.upload_file(path_or_fileobj=output_path, path_in_repo=f"{eval_request.model}/results_{datetime.now()}.json",
repo_id=RESULTS_REPO, repo_type="dataset")
set_eval_request(api=API, eval_request=eval_request, set_to_status=FINISHED_STATUS, hf_repo=QUEUE_REPO,
local_dir=EVAL_REQUESTS_PATH_BACKEND)
# breakpoint()
if __name__ == "__main__":
run_auto_eval()