Spaces:
Runtime error
Runtime error
Commit ·
a5c5307
1
Parent(s): 2b56c10
switched to child processes
Browse files- src/envs.py +1 -1
- src/leaderboard/build_leaderboard.py +4 -5
src/envs.py
CHANGED
|
@@ -15,7 +15,7 @@ PRIVATE_RESULTS_REPO = "open-llm-leaderboard/private-results"
|
|
| 15 |
|
| 16 |
IS_PUBLIC = bool(os.environ.get("IS_PUBLIC", True))
|
| 17 |
|
| 18 |
-
HF_HOME = os.getenv("HF_HOME", "
|
| 19 |
|
| 20 |
# Check HF_HOME write access
|
| 21 |
print(f"Initial HF_HOME set to: {HF_HOME}")
|
|
|
|
| 15 |
|
| 16 |
IS_PUBLIC = bool(os.environ.get("IS_PUBLIC", True))
|
| 17 |
|
| 18 |
+
HF_HOME = os.getenv("HF_HOME", ".")
|
| 19 |
|
| 20 |
# Check HF_HOME write access
|
| 21 |
print(f"Initial HF_HOME set to: {HF_HOME}")
|
src/leaderboard/build_leaderboard.py
CHANGED
|
@@ -61,10 +61,11 @@ def download_openbench():
|
|
| 61 |
# download answers of different models that we trust
|
| 62 |
download_dataset("Vikhrmodels/openbench-eval", EVAL_RESULTS_PATH)
|
| 63 |
|
| 64 |
-
logging.info(subprocess.
|
|
|
|
| 65 |
|
| 66 |
# copy the trusted model answers to data
|
| 67 |
-
subprocess.
|
| 68 |
[
|
| 69 |
"rsync",
|
| 70 |
"-azP",
|
|
@@ -72,11 +73,10 @@ def download_openbench():
|
|
| 72 |
f"{EVAL_RESULTS_PATH}/internal/*",
|
| 73 |
"data/arena-hard-v0.1/model_answer/internal/",
|
| 74 |
],
|
| 75 |
-
check=False,
|
| 76 |
)
|
| 77 |
# copy the judgement pre generated
|
| 78 |
# Will be rewritten after we switch to new gen for each submit
|
| 79 |
-
subprocess.
|
| 80 |
[
|
| 81 |
"rsync",
|
| 82 |
"-azP",
|
|
@@ -84,7 +84,6 @@ def download_openbench():
|
|
| 84 |
f"{EVAL_RESULTS_PATH}/model_judgment/*",
|
| 85 |
"data/arena-hard-v0.1/model_judgement/",
|
| 86 |
],
|
| 87 |
-
check=False,
|
| 88 |
)
|
| 89 |
|
| 90 |
|
|
|
|
| 61 |
# download answers of different models that we trust
|
| 62 |
download_dataset("Vikhrmodels/openbench-eval", EVAL_RESULTS_PATH)
|
| 63 |
|
| 64 |
+
logging.info(subprocess.Popen("pwd"))
|
| 65 |
+
logging.info(subprocess.Popen(f"ls {EVAL_RESULTS_PATH}"))
|
| 66 |
|
| 67 |
# copy the trusted model answers to data
|
| 68 |
+
subprocess.Popen(
|
| 69 |
[
|
| 70 |
"rsync",
|
| 71 |
"-azP",
|
|
|
|
| 73 |
f"{EVAL_RESULTS_PATH}/internal/*",
|
| 74 |
"data/arena-hard-v0.1/model_answer/internal/",
|
| 75 |
],
|
|
|
|
| 76 |
)
|
| 77 |
# copy the judgement pre generated
|
| 78 |
# Will be rewritten after we switch to new gen for each submit
|
| 79 |
+
subprocess.Popen(
|
| 80 |
[
|
| 81 |
"rsync",
|
| 82 |
"-azP",
|
|
|
|
| 84 |
f"{EVAL_RESULTS_PATH}/model_judgment/*",
|
| 85 |
"data/arena-hard-v0.1/model_judgement/",
|
| 86 |
],
|
|
|
|
| 87 |
)
|
| 88 |
|
| 89 |
|