Spaces:
Sleeping
Sleeping
File size: 3,692 Bytes
460d762 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 |
from dataclasses import dataclass
import glob
import json
from typing import Dict, List, Tuple
from src.utils_display import AutoEvalColumn, make_clickable_model
import numpy as np
# clone / pull the lmeh eval data
METRICS = ["acc_norm", "acc_norm", "acc_norm", "mc2"]
BENCHMARKS = ["arc_challenge", "hellaswag", "hendrycks", "truthfulqa_mc"]
BENCH_TO_NAME = {
"arc_challenge": AutoEvalColumn.arc.name,
"hellaswag": AutoEvalColumn.hellaswag.name,
"hendrycks": AutoEvalColumn.mmlu.name,
"truthfulqa_mc": AutoEvalColumn.truthfulqa.name,
}
@dataclass
class EvalResult:
eval_name: str
org: str
model: str
revision: str
is_8bit: bool
results: dict
def to_dict(self):
if self.org is not None:
base_model = f"{self.org}/{self.model}"
else:
base_model = f"{self.model}"
data_dict = {}
data_dict["eval_name"] = self.eval_name # not a column, just a save name
data_dict[AutoEvalColumn.is_8bit.name] = self.is_8bit
data_dict[AutoEvalColumn.model.name] = make_clickable_model(base_model)
data_dict[AutoEvalColumn.dummy.name] = base_model
data_dict[AutoEvalColumn.revision.name] = self.revision
data_dict[AutoEvalColumn.average.name] = round(
sum([v for k, v in self.results.items()]) / 4.0, 1
)
for benchmark in BENCHMARKS:
if not benchmark in self.results.keys():
self.results[benchmark] = None
for k, v in BENCH_TO_NAME.items():
data_dict[v] = self.results[k]
return data_dict
def parse_eval_result(json_filepath: str) -> Tuple[str, dict]:
with open(json_filepath) as fp:
data = json.load(fp)
path_split = json_filepath.split("/")
org = None
model = path_split[-4]
is_8bit = path_split[-2] == "8bit"
revision = path_split[-3]
if len(path_split) == 7:
# handles gpt2 type models that don't have an org
result_key = f"{model}_{revision}_{is_8bit}"
else:
org = path_split[-5]
result_key = f"{org}_{model}_{revision}_{is_8bit}"
eval_result = None
for benchmark, metric in zip(BENCHMARKS, METRICS):
if benchmark in json_filepath:
accs = np.array([v[metric] for v in data["results"].values()])
mean_acc = round(np.mean(accs) * 100.0, 1)
eval_result = EvalResult(
result_key, org, model, revision, is_8bit, {benchmark: mean_acc}
)
return result_key, eval_result
def get_eval_results(is_public) -> List[EvalResult]:
json_filepaths = glob.glob(
"auto_evals/eval_results/public/**/16bit/*.json", recursive=True
)
if not is_public:
json_filepaths += glob.glob(
"auto_evals/eval_results/private/**/*.json", recursive=True
)
json_filepaths += glob.glob(
"auto_evals/eval_results/private/**/*.json", recursive=True
)
# include the 8bit evals of public models
json_filepaths += glob.glob(
"auto_evals/eval_results/public/**/8bit/*.json", recursive=True
)
eval_results = {}
for json_filepath in json_filepaths:
result_key, eval_result = parse_eval_result(json_filepath)
if result_key in eval_results.keys():
eval_results[result_key].results.update(eval_result.results)
else:
eval_results[result_key] = eval_result
eval_results = [v for v in eval_results.values()]
return eval_results
def get_eval_results_dicts(is_public=True) -> List[Dict]:
eval_results = get_eval_results(is_public)
return [e.to_dict() for e in eval_results]
|