tohid.abedini
commited on
Commit
·
00e3432
1
Parent(s):
ac38bfe
[Add] fake results jsonl
Browse files- app.py +2 -1
- leaderboard_base.json +12 -36
- leaderboard_persian.json +12 -36
- leaderboard_persian.jsonl +28 -0
app.py
CHANGED
|
@@ -100,7 +100,8 @@ def submit(model_name, model_id, contact_email, challenge, submission_id, paper_
|
|
| 100 |
abs_path = Path(__file__).parent
|
| 101 |
|
| 102 |
# Any pandas-compatible data
|
| 103 |
-
persian_df = pd.read_json(str(abs_path / "leaderboard_persian.
|
|
|
|
| 104 |
base_df = pd.read_json(str(abs_path / "leaderboard_base.json"))
|
| 105 |
|
| 106 |
with gr.Blocks(css=custom_css) as demo:
|
|
|
|
| 100 |
abs_path = Path(__file__).parent
|
| 101 |
|
| 102 |
# Any pandas-compatible data
|
| 103 |
+
# persian_df = pd.read_json(str(abs_path / "leaderboard_persian.jsonl"))
|
| 104 |
+
persian_df = pd.read_json(str(abs_path / "leaderboard_persian.jsonl"), lines=True)
|
| 105 |
base_df = pd.read_json(str(abs_path / "leaderboard_base.json"))
|
| 106 |
|
| 107 |
with gr.Blocks(css=custom_css) as demo:
|
leaderboard_base.json
CHANGED
|
@@ -1,38 +1,14 @@
|
|
| 1 |
{
|
| 2 |
-
"Model": {
|
| 3 |
-
|
| 4 |
-
},
|
| 5 |
-
"
|
| 6 |
-
|
| 7 |
-
},
|
| 8 |
-
"
|
| 9 |
-
|
| 10 |
-
},
|
| 11 |
-
"
|
| 12 |
-
|
| 13 |
-
}
|
| 14 |
-
"ARC Easy": {
|
| 15 |
-
"0": 65
|
| 16 |
-
},
|
| 17 |
-
"ARC Challenging": {
|
| 18 |
-
"0": 45
|
| 19 |
-
},
|
| 20 |
-
"MMLU Pro": {
|
| 21 |
-
"0": 35
|
| 22 |
-
},
|
| 23 |
-
"GSM8k Persian": {
|
| 24 |
-
"0": 25
|
| 25 |
-
},
|
| 26 |
-
"Multiple Choice Persian": {
|
| 27 |
-
"0": 15
|
| 28 |
-
},
|
| 29 |
-
"Hub License": {
|
| 30 |
-
"0": "llama3"
|
| 31 |
-
},
|
| 32 |
-
"Model sha": {
|
| 33 |
-
"0": "5f10e376dcdac39811d4ad43424f6122635784a3"
|
| 34 |
-
},
|
| 35 |
-
"model_name_for_query": {
|
| 36 |
-
"0": "PartAI/Dorna-Llama3-8B-Instruct"
|
| 37 |
-
}
|
| 38 |
}
|
|
|
|
| 1 |
{
|
| 2 |
+
"Model": {"0": "meta-llama/Meta-Llama-3-8B-Instruct"},
|
| 3 |
+
"Precision": {"0": "bfloat16"},
|
| 4 |
+
"#Params (B)": {"0": 8.03},
|
| 5 |
+
"Part Multiple Choice": {"0": 54},
|
| 6 |
+
"ARC Easy": {"0": 64},
|
| 7 |
+
"ARC Challenging": {"0": 44},
|
| 8 |
+
"MMLU Pro": {"0": 34},
|
| 9 |
+
"GSM8k Persian": {"0": 15},
|
| 10 |
+
"Multiple Choice Persian": {"0": 5},
|
| 11 |
+
"Hub License": {"0": "llama3"},
|
| 12 |
+
"Model sha": {"0": "5f10e376dcdac39811d4ad43424f6122635784a3"},
|
| 13 |
+
"model_name_for_query": {"0": "meta-llama/Meta-Llama-3-8B-Instruct"}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
}
|
leaderboard_persian.json
CHANGED
|
@@ -1,38 +1,14 @@
|
|
| 1 |
{
|
| 2 |
-
"Model": {
|
| 3 |
-
|
| 4 |
-
},
|
| 5 |
-
"
|
| 6 |
-
|
| 7 |
-
},
|
| 8 |
-
"
|
| 9 |
-
|
| 10 |
-
},
|
| 11 |
-
"
|
| 12 |
-
|
| 13 |
-
}
|
| 14 |
-
"ARC Easy": {
|
| 15 |
-
"0": 64
|
| 16 |
-
},
|
| 17 |
-
"ARC Challenging": {
|
| 18 |
-
"0": 44
|
| 19 |
-
},
|
| 20 |
-
"MMLU Pro": {
|
| 21 |
-
"0": 34
|
| 22 |
-
},
|
| 23 |
-
"GSM8k Persian": {
|
| 24 |
-
"0": 15
|
| 25 |
-
},
|
| 26 |
-
"Multiple Choice Persian": {
|
| 27 |
-
"0": 5
|
| 28 |
-
},
|
| 29 |
-
"Hub License": {
|
| 30 |
-
"0": "llama3"
|
| 31 |
-
},
|
| 32 |
-
"Model sha": {
|
| 33 |
-
"0": "5f10e376dcdac39811d4ad43424f6122635784a3"
|
| 34 |
-
},
|
| 35 |
-
"model_name_for_query": {
|
| 36 |
-
"0": "meta-llama/Meta-Llama-3-8B-Instruct"
|
| 37 |
-
}
|
| 38 |
}
|
|
|
|
| 1 |
{
|
| 2 |
+
"Model": {"0": "PartAI/Dorna-Llama3-8B-Instruct"},
|
| 3 |
+
"Precision": {"0": "bfloat16"},
|
| 4 |
+
"#Params (B)": {"0": 8.03},
|
| 5 |
+
"Part Multiple Choice": {"0": 55},
|
| 6 |
+
"ARC Easy": {"0": 65},
|
| 7 |
+
"ARC Challenging": {"0": 45},
|
| 8 |
+
"MMLU Pro": {"0": 35},
|
| 9 |
+
"GSM8k Persian": {"0": 25},
|
| 10 |
+
"Multiple Choice Persian": {"0": 15},
|
| 11 |
+
"Hub License": {"0": "llama3"},
|
| 12 |
+
"Model sha": {"0": "5f10e376dcdac39811d4ad43424f6122635784a3"},
|
| 13 |
+
"model_name_for_query": {"0": "PartAI/Dorna-Llama3-8B-Instruct"}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
}
|
leaderboard_persian.jsonl
ADDED
|
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"Model": "PartAI/Dorna-Llama3-8B-Instruct",
|
| 3 |
+
"Precision": "bfloat16",
|
| 4 |
+
"#Params (B)": 8.03,
|
| 5 |
+
"Part Multiple Choice": 55,
|
| 6 |
+
"ARC Easy": 65,
|
| 7 |
+
"ARC Challenging": 45,
|
| 8 |
+
"MMLU Pro": 35,
|
| 9 |
+
"GSM8k Persian": 25,
|
| 10 |
+
"Multiple Choice Persian": 15,
|
| 11 |
+
"Hub License": "llama3",
|
| 12 |
+
"Model sha": "5f10e376dcdac39811d4ad43424f6122635784a3",
|
| 13 |
+
"model_name_for_query": "PartAI/Dorna-Llama3-8B-Instruct"
|
| 14 |
+
}
|
| 15 |
+
{
|
| 16 |
+
"Model": "PartAI/Dorna-Llama3-35B-Instruct",
|
| 17 |
+
"Precision": "bfloat16",
|
| 18 |
+
"#Params (B)": 35.03,
|
| 19 |
+
"Part Multiple Choice": 55,
|
| 20 |
+
"ARC Easy": 65,
|
| 21 |
+
"ARC Challenging": 45,
|
| 22 |
+
"MMLU Pro": 35,
|
| 23 |
+
"GSM8k Persian": 25,
|
| 24 |
+
"Multiple Choice Persian": 15,
|
| 25 |
+
"Hub License": "llama3",
|
| 26 |
+
"Model sha": "5f10e376dcdac39811d4ad43424f6122635784a3",
|
| 27 |
+
"model_name_for_query": "PartAI/Dorna-Llama3-35B-Instruct"
|
| 28 |
+
}
|