Revise cybergym
Browse files- gen_table.py +1 -1
gen_table.py
CHANGED
|
@@ -56,7 +56,7 @@ def BUILD_L2_DF(results, benchmark):
|
|
| 56 |
res = defaultdict(list)
|
| 57 |
if benchmark not in ["RedCode","NYU CTF Bench","PrimeVul","SWE-bench-verified"]:
|
| 58 |
res['Model']=model_list
|
| 59 |
-
elif benchmark=="SWE-bench-verified":
|
| 60 |
res['Agent']=model_list
|
| 61 |
elif benchmark == "PrimeVul":
|
| 62 |
used=[]
|
|
|
|
| 56 |
res = defaultdict(list)
|
| 57 |
if benchmark not in ["RedCode","NYU CTF Bench","PrimeVul","SWE-bench-verified"]:
|
| 58 |
res['Model']=model_list
|
| 59 |
+
elif benchmark=="SWE-bench-verified" or benchmark=="CyberGym":
|
| 60 |
res['Agent']=model_list
|
| 61 |
elif benchmark == "PrimeVul":
|
| 62 |
used=[]
|