29 lines
656 B
JSON
29 lines
656 B
JSON
|
{
|
||
|
|
"results": {
|
||
|
|
"hellaswag": {
|
||
|
|
"acc": 0.5736904999004182,
|
||
|
|
"acc_stderr": 0.004935291975579188,
|
||
|
|
"acc_norm": 0.7689703246365266,
|
||
|
|
"acc_norm_stderr": 0.004206299481687387
|
||
|
|
}
|
||
|
|
},
|
||
|
|
"versions": {
|
||
|
|
"hellaswag": 0
|
||
|
|
},
|
||
|
|
"config": {
|
||
|
|
"model": "hf-causal-experimental",
|
||
|
|
"model_args": {
|
||
|
|
"pretrained": "/network/alexandre/research/cerebras/retrained_s50",
|
||
|
|
"trust_remote_code": true,
|
||
|
|
"dtype": "bfloat16"
|
||
|
|
},
|
||
|
|
"num_fewshot": 10,
|
||
|
|
"batch_size": "8",
|
||
|
|
"batch_sizes": [],
|
||
|
|
"device": "cuda:2",
|
||
|
|
"no_cache": true,
|
||
|
|
"limit": null,
|
||
|
|
"bootstrap_iters": 100000,
|
||
|
|
"description_dict": {}
|
||
|
|
}
|
||
|
|
}
|