metadata fixed
Browse files- functions.py +3 -3
functions.py
CHANGED
@@ -41,7 +41,7 @@ def get_task_summary(results):
|
|
41 |
"IFEval":
|
42 |
{"dataset_type":"HuggingFaceH4/ifeval",
|
43 |
"dataset_name":"IFEval (0-Shot)",
|
44 |
-
"metric_type":"inst_level_strict_acc",
|
45 |
"metric_value":results["IFEval"],
|
46 |
"dataset_config": None, # don't know
|
47 |
"dataset_split": None, # don't know
|
@@ -84,10 +84,10 @@ def get_task_summary(results):
|
|
84 |
"dataset_args":{"num_few_shot": 0},
|
85 |
"metric_name":"acc_norm"
|
86 |
},
|
87 |
-
"
|
88 |
{
|
89 |
"dataset_type":"TAUR-Lab/MuSR",
|
90 |
-
"dataset_name":"
|
91 |
"metric_type":"acc_norm",
|
92 |
"metric_value":results["MUSR"],
|
93 |
"dataset_config": None, # don't know
|
|
|
41 |
"IFEval":
|
42 |
{"dataset_type":"HuggingFaceH4/ifeval",
|
43 |
"dataset_name":"IFEval (0-Shot)",
|
44 |
+
"metric_type": ["inst_level_strict_acc", "prompt_level_strict_acc"],
|
45 |
"metric_value":results["IFEval"],
|
46 |
"dataset_config": None, # don't know
|
47 |
"dataset_split": None, # don't know
|
|
|
84 |
"dataset_args":{"num_few_shot": 0},
|
85 |
"metric_name":"acc_norm"
|
86 |
},
|
87 |
+
"MuSR":
|
88 |
{
|
89 |
"dataset_type":"TAUR-Lab/MuSR",
|
90 |
+
"dataset_name":"MuSR (0-shot)",
|
91 |
"metric_type":"acc_norm",
|
92 |
"metric_value":results["MUSR"],
|
93 |
"dataset_config": None, # don't know
|