Spaces:
Runtime error
Runtime error
File size: 3,929 Bytes
460d762 4aff44e 460d762 d16cee2 460d762 d16cee2 460d762 d16cee2 460d762 12cea14 460d762 12cea14 460d762 d16cee2 460d762 d16cee2 460d762 4aff44e 97b27da 460d762 4aff44e d16cee2 460d762 d16cee2 460d762 d16cee2 460d762 d16cee2 460d762 d16cee2 460d762 d16cee2 460d762 d16cee2 460d762 d16cee2 460d762 d16cee2 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 |
from dataclasses import dataclass
import glob
import json
import os
from typing import Dict, List, Tuple
from src.utils_display import AutoEvalColumn, make_clickable_model
import numpy as np
METRICS = ["acc_norm", "acc_norm", "acc", "mc2"]
BENCHMARKS = ["arc:challenge", "hellaswag", "hendrycksTest", "truthfulqa:mc"]
BENCH_TO_NAME = {
"arc:challenge": AutoEvalColumn.arc.name,
"hellaswag": AutoEvalColumn.hellaswag.name,
"hendrycksTest": AutoEvalColumn.mmlu.name,
"truthfulqa:mc": AutoEvalColumn.truthfulqa.name,
}
@dataclass
class EvalResult:
eval_name: str
org: str
model: str
revision: str
results: dict
precision: str = "16bit"
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.precision.name] = self.precision
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 benchmark not 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, list[dict]]:
with open(json_filepath) as fp:
data = json.load(fp)
for mmlu_k in ["harness|hendrycksTest-abstract_algebra|5", "hendrycksTest-abstract_algebra"]:
if mmlu_k in data["versions"] and data["versions"][mmlu_k] == 0:
return None, [] # we skip models with the wrong version
try:
config = data["config"]
except KeyError:
config = data["config_general"]
model = config.get("model_name", None)
if model is None:
model = config.get("model_args", None)
model_sha = config.get("model_sha", "")
eval_sha = config.get("lighteval_sha", "")
model_split = model.split("/", 1)
model = model_split[-1]
if len(model_split) == 1:
org = None
model = model_split[0]
result_key = f"{model}_{model_sha}_{eval_sha}"
else:
org = model_split[0]
model = model_split[1]
result_key = f"{org}_{model}_{model_sha}_{eval_sha}"
eval_results = []
for benchmark, metric in zip(BENCHMARKS, METRICS):
accs = np.array([v[metric] for k, v in data["results"].items() if benchmark in k])
if accs.size == 0:
continue
mean_acc = round(np.mean(accs) * 100.0, 1)
eval_results.append(EvalResult(
result_key, org, model, model_sha, {benchmark: mean_acc}
))
return result_key, eval_results
def get_eval_results(is_public) -> List[EvalResult]:
json_filepaths = glob.glob(
"eval-results/**/results*.json", recursive=True
)
if not is_public:
json_filepaths += glob.glob(
"private-eval-results/**/results*.json", recursive=True
)
eval_results = {}
for json_filepath in json_filepaths:
result_key, results = parse_eval_result(json_filepath)
for eval_result in results:
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]
|