Muennighoff
commited on
Commit
·
26444b6
1
Parent(s):
112233b
Add
Browse files- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed1_0_babi.json +22 -0
- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed1_1_babi.json +22 -0
- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed1_2_babi.json +22 -0
- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed1_3_babi.json +22 -0
- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed1_4_babi.json +22 -0
- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed1_5_babi.json +22 -0
- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed2_0_babi.json +22 -0
- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed2_1_babi.json +22 -0
- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed2_2_babi.json +22 -0
- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed2_3_babi.json +22 -0
- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed2_4_babi.json +22 -0
- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed2_5_babi.json +22 -0
- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed3_0_babi.json +22 -0
- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed3_1_babi.json +22 -0
- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed3_2_babi.json +22 -0
- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed3_3_babi.json +22 -0
- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed3_4_babi.json +22 -0
- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed3_5_babi.json +22 -0
- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed4_0_babi.json +22 -0
- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed4_1_babi.json +22 -0
- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed4_2_babi.json +22 -0
- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed4_3_babi.json +22 -0
- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed4_4_babi.json +22 -0
- 4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed4_5_babi.json +22 -0
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed1_0_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed1/transformers",
|
14 |
+
"num_fewshot": 0,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed1_1_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed1/transformers",
|
14 |
+
"num_fewshot": 1,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed1_2_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed1/transformers",
|
14 |
+
"num_fewshot": 2,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed1_3_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed1/transformers",
|
14 |
+
"num_fewshot": 3,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed1_4_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed1/transformers",
|
14 |
+
"num_fewshot": 4,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed1_5_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed1/transformers",
|
14 |
+
"num_fewshot": 5,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed2_0_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed2/transformers",
|
14 |
+
"num_fewshot": 0,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed2_1_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed2/transformers",
|
14 |
+
"num_fewshot": 1,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed2_2_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed2/transformers",
|
14 |
+
"num_fewshot": 2,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed2_3_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed2/transformers",
|
14 |
+
"num_fewshot": 3,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed2_4_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed2/transformers",
|
14 |
+
"num_fewshot": 4,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed2_5_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed2/transformers",
|
14 |
+
"num_fewshot": 5,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed3_0_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed3/transformers",
|
14 |
+
"num_fewshot": 0,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed3_1_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed3/transformers",
|
14 |
+
"num_fewshot": 1,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed3_2_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed3/transformers",
|
14 |
+
"num_fewshot": 2,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed3_3_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed3/transformers",
|
14 |
+
"num_fewshot": 3,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed3_4_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed3/transformers",
|
14 |
+
"num_fewshot": 4,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed3_5_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed3/transformers",
|
14 |
+
"num_fewshot": 5,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed4_0_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed4/transformers",
|
14 |
+
"num_fewshot": 0,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed4_1_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed4/transformers",
|
14 |
+
"num_fewshot": 1,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed4_2_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed4/transformers",
|
14 |
+
"num_fewshot": 2,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed4_3_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed4/transformers",
|
14 |
+
"num_fewshot": 3,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed4_4_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed4/transformers",
|
14 |
+
"num_fewshot": 4,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|
4b284b84bc4v2seed1/evaluation/4b284b84bc4v2seed4_5_babi.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"babi": {
|
4 |
+
"em": 0.0,
|
5 |
+
"em_stderr": 0.0
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"babi": 0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "gpt2",
|
13 |
+
"model_args": "pretrained=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4v2seed4/transformers",
|
14 |
+
"num_fewshot": 5,
|
15 |
+
"batch_size": null,
|
16 |
+
"device": null,
|
17 |
+
"no_cache": true,
|
18 |
+
"limit": 3000,
|
19 |
+
"bootstrap_iters": 100000,
|
20 |
+
"description_dict": {}
|
21 |
+
}
|
22 |
+
}
|