Anas Awadalla
commited on
Commit
·
dec2f9b
1
Parent(s):
fe4ed4c
osworld baselines
Browse files- src/streamlit_app.py +15 -0
src/streamlit_app.py
CHANGED
@@ -76,9 +76,24 @@ BASELINES = {
|
|
76 |
"UI-TARS-1.5-7B": {"overall": 67.2},
|
77 |
},
|
78 |
"osworld-g": {
|
|
|
|
|
|
|
79 |
"Qwen2.5VL-3B": {"overall": 27.3},
|
|
|
80 |
"Qwen2.5VL-7B": {"overall": 31.4},
|
|
|
|
|
81 |
"UI-TARS-7B": {"overall": 47.5},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
82 |
}
|
83 |
}
|
84 |
|
|
|
76 |
"UI-TARS-1.5-7B": {"overall": 67.2},
|
77 |
},
|
78 |
"osworld-g": {
|
79 |
+
"Operator": {"overall": 40.6},
|
80 |
+
"Gemini-2.5-Pro": {"overall": 45.2},
|
81 |
+
"Seed1.5-VL": {"overall": 62.9},
|
82 |
"Qwen2.5VL-3B": {"overall": 27.3},
|
83 |
+
"OS-Atlas-7B": {"overall": 27.7},
|
84 |
"Qwen2.5VL-7B": {"overall": 31.4},
|
85 |
+
"UGround-7B": {"overall": 36.4},
|
86 |
+
"Aguvis-7B": {"overall": 38.7},
|
87 |
"UI-TARS-7B": {"overall": 47.5},
|
88 |
+
"Qwen2.5-VL-32B*": {"overall": 59.6},
|
89 |
+
"Jedi-3B": {"overall": 50.9},
|
90 |
+
"Jedi-7B": {"overall": 54.1},
|
91 |
+
"UI-TARS-72B": {"overall": 57.1},
|
92 |
+
"Qwen2.5-VL-72B*": {"overall": 62.2},
|
93 |
+
"UI-TARS-1.5-7B*": {"overall": 64.2},
|
94 |
+
"GTAI-7B": {"overall": 67.7},
|
95 |
+
"GTAI-32B": {"overall": 61.9},
|
96 |
+
"GTAI-72B": {"overall": 66.7},
|
97 |
}
|
98 |
}
|
99 |
|