Bram Vanroy
commited on
Commit
·
58ced9f
1
Parent(s):
4896850
remove old ultra
Browse files
evals/arc/arc_nl_GEITje-ultra.json
DELETED
|
@@ -1,23 +0,0 @@
|
|
| 1 |
-
{
|
| 2 |
-
"results": {
|
| 3 |
-
"arc_nl": {
|
| 4 |
-
"acc": 0.40633019674935844,
|
| 5 |
-
"acc_stderr": 0.01437111972522202,
|
| 6 |
-
"acc_norm": 0.41060735671514115,
|
| 7 |
-
"acc_norm_stderr": 0.014394424200139139
|
| 8 |
-
}
|
| 9 |
-
},
|
| 10 |
-
"versions": {
|
| 11 |
-
"arc_nl": 0
|
| 12 |
-
},
|
| 13 |
-
"config": {
|
| 14 |
-
"model": "hf-auto",
|
| 15 |
-
"model_args": "pretrained=BramVanroy/GEITje-ultra-dpo-1e-07lr-0.01b-1ep-128tbs,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-ultra.json
DELETED
|
@@ -1,23 +0,0 @@
|
|
| 1 |
-
{
|
| 2 |
-
"results": {
|
| 3 |
-
"hellaswag_nl": {
|
| 4 |
-
"acc": 0.535887749595251,
|
| 5 |
-
"acc_stderr": 0.005181424297774049,
|
| 6 |
-
"acc_norm": 0.6670264436049649,
|
| 7 |
-
"acc_norm_stderr": 0.0048964026564578735
|
| 8 |
-
}
|
| 9 |
-
},
|
| 10 |
-
"versions": {
|
| 11 |
-
"hellaswag_nl": 1
|
| 12 |
-
},
|
| 13 |
-
"config": {
|
| 14 |
-
"model": "hf-auto",
|
| 15 |
-
"model_args": "pretrained=BramVanroy/GEITje-ultra-dpo-1e-07lr-0.01b-1ep-128tbs,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-ultra.json
DELETED
|
@@ -1,23 +0,0 @@
|
|
| 1 |
-
{
|
| 2 |
-
"results": {
|
| 3 |
-
"mmlu_nl": {
|
| 4 |
-
"acc": 0.4178492828413144,
|
| 5 |
-
"acc_stderr": 0.004296707163181694,
|
| 6 |
-
"acc_norm": 0.35698565682628824,
|
| 7 |
-
"acc_norm_stderr": 0.004173918076627818
|
| 8 |
-
}
|
| 9 |
-
},
|
| 10 |
-
"versions": {
|
| 11 |
-
"mmlu_nl": 0
|
| 12 |
-
},
|
| 13 |
-
"config": {
|
| 14 |
-
"model": "hf-auto",
|
| 15 |
-
"model_args": "pretrained=BramVanroy/GEITje-ultra-dpo-1e-07lr-0.01b-1ep-128tbs,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-ultra.json
DELETED
|
@@ -1,23 +0,0 @@
|
|
| 1 |
-
{
|
| 2 |
-
"results": {
|
| 3 |
-
"truthfulqa_nl": {
|
| 4 |
-
"mc1": 0.3235668789808917,
|
| 5 |
-
"mc1_stderr": 0.016708460287009546,
|
| 6 |
-
"mc2": 0.5186509420055189,
|
| 7 |
-
"mc2_stderr": 0.01627138410084977
|
| 8 |
-
}
|
| 9 |
-
},
|
| 10 |
-
"versions": {
|
| 11 |
-
"truthfulqa_nl": 1
|
| 12 |
-
},
|
| 13 |
-
"config": {
|
| 14 |
-
"model": "hf-auto",
|
| 15 |
-
"model_args": "pretrained=BramVanroy/GEITje-ultra-dpo-1e-07lr-0.01b-1ep-128tbs,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 |
-
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|