Spaces:
Sleeping
Sleeping
Commit
•
524e209
1
Parent(s):
a4b20f4
Remove ARC task
Browse files- src/constants.py +2 -2
src/constants.py
CHANGED
@@ -14,7 +14,7 @@ RESULTS_DATASET_ID = "datasets/open-llm-leaderboard/results"
|
|
14 |
DETAILS_DATASET_ID = "datasets/open-llm-leaderboard/{model_name_sanitized}-details"
|
15 |
DETAILS_FILENAME = "samples_{subtask}_*.json"
|
16 |
TASKS = {
|
17 |
-
"leaderboard_arc_challenge": ("ARC", "leaderboard_arc_challenge"),
|
18 |
"leaderboard_bbh": ("BBH", "leaderboard_bbh"),
|
19 |
"leaderboard_gpqa": ("GPQA", "leaderboard_gpqa"),
|
20 |
"leaderboard_ifeval": ("IFEval", "leaderboard_ifeval"),
|
@@ -23,7 +23,7 @@ TASKS = {
|
|
23 |
"leaderboard_musr": ("MuSR", "leaderboard_musr"),
|
24 |
}
|
25 |
SUBTASKS = {
|
26 |
-
"leaderboard_arc_challenge": ["leaderboard_arc_challenge"],
|
27 |
"leaderboard_bbh": [
|
28 |
"leaderboard_bbh_boolean_expressions",
|
29 |
"leaderboard_bbh_causal_judgement",
|
|
|
14 |
DETAILS_DATASET_ID = "datasets/open-llm-leaderboard/{model_name_sanitized}-details"
|
15 |
DETAILS_FILENAME = "samples_{subtask}_*.json"
|
16 |
TASKS = {
|
17 |
+
# "leaderboard_arc_challenge": ("ARC", "leaderboard_arc_challenge"),
|
18 |
"leaderboard_bbh": ("BBH", "leaderboard_bbh"),
|
19 |
"leaderboard_gpqa": ("GPQA", "leaderboard_gpqa"),
|
20 |
"leaderboard_ifeval": ("IFEval", "leaderboard_ifeval"),
|
|
|
23 |
"leaderboard_musr": ("MuSR", "leaderboard_musr"),
|
24 |
}
|
25 |
SUBTASKS = {
|
26 |
+
# "leaderboard_arc_challenge": ["leaderboard_arc_challenge"],
|
27 |
"leaderboard_bbh": [
|
28 |
"leaderboard_bbh_boolean_expressions",
|
29 |
"leaderboard_bbh_causal_judgement",
|