Bram Vanroy commited on
Commit ·
9928d85
1
Parent(s): f19fc85
remove old runs
Browse files
evals/arc/arc_nl_GEITje-7b-chat-v2.json
DELETED
|
@@ -1,23 +0,0 @@
|
|
| 1 |
-
{
|
| 2 |
-
"results": {
|
| 3 |
-
"arc_nl": {
|
| 4 |
-
"acc": 0.41916167664670656,
|
| 5 |
-
"acc_stderr": 0.01443766613132885,
|
| 6 |
-
"acc_norm": 0.42001710863986313,
|
| 7 |
-
"acc_norm_stderr": 0.014441744604939782
|
| 8 |
-
}
|
| 9 |
-
},
|
| 10 |
-
"versions": {
|
| 11 |
-
"arc_nl": 0
|
| 12 |
-
},
|
| 13 |
-
"config": {
|
| 14 |
-
"model": "hf-auto",
|
| 15 |
-
"model_args": "pretrained=Rijgersberg/GEITje-7b-chat-v2,use_accelerate=True,device_map_option=auto,dtype=auto,trust_remote_code=True",
|
| 16 |
-
"batch_size": 2,
|
| 17 |
-
"device": "cuda",
|
| 18 |
-
"no_cache": false,
|
| 19 |
-
"limit": null,
|
| 20 |
-
"bootstrap_iters": 100000,
|
| 21 |
-
"description_dict": {}
|
| 22 |
-
}
|
| 23 |
-
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
evals/hellaswag/hellaswag_nl_GEITje-7b-chat-v2.json
DELETED
|
@@ -1,23 +0,0 @@
|
|
| 1 |
-
{
|
| 2 |
-
"results": {
|
| 3 |
-
"hellaswag_nl": {
|
| 4 |
-
"acc": 0.5027522935779817,
|
| 5 |
-
"acc_stderr": 0.005194744014864852,
|
| 6 |
-
"acc_norm": 0.6514840798704803,
|
| 7 |
-
"acc_norm_stderr": 0.004950669577821165
|
| 8 |
-
}
|
| 9 |
-
},
|
| 10 |
-
"versions": {
|
| 11 |
-
"hellaswag_nl": 1
|
| 12 |
-
},
|
| 13 |
-
"config": {
|
| 14 |
-
"model": "hf-auto",
|
| 15 |
-
"model_args": "pretrained=Rijgersberg/GEITje-7b-chat-v2,use_accelerate=True,device_map_option=auto,dtype=auto,trust_remote_code=True",
|
| 16 |
-
"batch_size": 2,
|
| 17 |
-
"device": "cuda",
|
| 18 |
-
"no_cache": false,
|
| 19 |
-
"limit": null,
|
| 20 |
-
"bootstrap_iters": 100000,
|
| 21 |
-
"description_dict": {}
|
| 22 |
-
}
|
| 23 |
-
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
evals/mmlu/mmlu_nl_GEITje-7b-chat-v2.json
DELETED
|
@@ -1,23 +0,0 @@
|
|
| 1 |
-
{
|
| 2 |
-
"results": {
|
| 3 |
-
"mmlu_nl": {
|
| 4 |
-
"acc": 0.39250208696971994,
|
| 5 |
-
"acc_stderr": 0.004254040176630216,
|
| 6 |
-
"acc_norm": 0.33072778325870833,
|
| 7 |
-
"acc_norm_stderr": 0.004098688802134392
|
| 8 |
-
}
|
| 9 |
-
},
|
| 10 |
-
"versions": {
|
| 11 |
-
"mmlu_nl": 0
|
| 12 |
-
},
|
| 13 |
-
"config": {
|
| 14 |
-
"model": "hf-auto",
|
| 15 |
-
"model_args": "pretrained=Rijgersberg/GEITje-7b-chat-v2,use_accelerate=True,device_map_option=auto,dtype=auto,trust_remote_code=True",
|
| 16 |
-
"batch_size": 2,
|
| 17 |
-
"device": "cuda",
|
| 18 |
-
"no_cache": false,
|
| 19 |
-
"limit": null,
|
| 20 |
-
"bootstrap_iters": 100000,
|
| 21 |
-
"description_dict": {}
|
| 22 |
-
}
|
| 23 |
-
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
evals/truthfulqa/truthfulqa_nl_GEITje-7b-chat-v2.json
DELETED
|
@@ -1,23 +0,0 @@
|
|
| 1 |
-
{
|
| 2 |
-
"results": {
|
| 3 |
-
"truthfulqa_nl": {
|
| 4 |
-
"mc1": 0.2840764331210191,
|
| 5 |
-
"mc1_stderr": 0.016106193085169508,
|
| 6 |
-
"mc2": 0.4547595365833656,
|
| 7 |
-
"mc2_stderr": 0.015154288378174851
|
| 8 |
-
}
|
| 9 |
-
},
|
| 10 |
-
"versions": {
|
| 11 |
-
"truthfulqa_nl": 1
|
| 12 |
-
},
|
| 13 |
-
"config": {
|
| 14 |
-
"model": "hf-auto",
|
| 15 |
-
"model_args": "pretrained=Rijgersberg/GEITje-7b-chat-v2,use_accelerate=True,device_map_option=auto,dtype=auto,trust_remote_code=True",
|
| 16 |
-
"batch_size": 2,
|
| 17 |
-
"device": "cuda",
|
| 18 |
-
"no_cache": false,
|
| 19 |
-
"limit": null,
|
| 20 |
-
"bootstrap_iters": 100000,
|
| 21 |
-
"description_dict": {}
|
| 22 |
-
}
|
| 23 |
-
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|