pminervini's picture
update
dc1ba50
raw
history blame
637 Bytes
#!/usr/bin/env python
from huggingface_hub import snapshot_download
from src.leaderboard.read_evals import get_raw_eval_results
from src.envs import API, EVAL_REQUESTS_PATH, EVAL_RESULTS_PATH, QUEUE_REPO, RESULTS_REPO
snapshot_download(repo_id=QUEUE_REPO, local_dir=EVAL_REQUESTS_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30)
snapshot_download(repo_id=RESULTS_REPO, local_dir=EVAL_RESULTS_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30)
raw_data = get_raw_eval_results(EVAL_RESULTS_PATH, EVAL_REQUESTS_PATH)
for entry in raw_data:
if '125m' in entry.eval_name:
print(entry)
# print(raw_data)