File size: 4,961 Bytes
460d762 4aff44e 8c49cb6 460d762 8c49cb6 460d762 8c49cb6 d16cee2 460d762 d16cee2 460d762 d16cee2 460d762 6eaad72 b323764 6e79cea 5228101 460d762 8c49cb6 460d762 8c49cb6 b323764 12cea14 b323764 460d762 d350941 8c49cb6 460d762 d16cee2 460d762 d16cee2 460d762 4aff44e 97b27da 8c49cb6 460d762 4aff44e d16cee2 6eaad72 d16cee2 5228101 460d762 d16cee2 5228101 460d762 d16cee2 460d762 ba25d90 d16cee2 d350941 8c49cb6 5228101 8c49cb6 460d762 d16cee2 460d762 699e8ff d6b3d82 d16cee2 d6b3d82 6eaad72 d6b3d82 5228101 d6b3d82 6eaad72 d6b3d82 8c49cb6 6eaad72 d6b3d82 460d762 d16cee2 460d762 699e8ff 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 |
import json
import os
from dataclasses import dataclass
from typing import Dict, List, Tuple
import dateutil
import numpy as np
from src.display_models.utils import AutoEvalColumn, make_clickable_model
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 = ""
model_type: str = ""
weight_type: str = "Original"
date: str = ""
def to_dict(self):
from src.load_from_hub import is_model_on_hub
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["weight_type"] = self.weight_type # not a column, just a save name
data_dict[AutoEvalColumn.precision.name] = self.precision
data_dict[AutoEvalColumn.model_type.name] = self.model_type
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] = sum([v for k, v in self.results.items()]) / 4.0
data_dict[AutoEvalColumn.still_on_hub.name] = (
is_model_on_hub(base_model, self.revision)[0] or base_model == "baseline"
)
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", "")
model_split = model.split("/", 1)
precision = config.get("model_dtype")
model = model_split[-1]
if len(model_split) == 1:
org = None
model = model_split[0]
result_key = f"{model}_{precision}"
else:
org = model_split[0]
model = model_split[1]
result_key = f"{org}_{model}_{precision}"
eval_results = []
for benchmark, metric in zip(BENCHMARKS, METRICS):
accs = np.array([v.get(metric, None) for k, v in data["results"].items() if benchmark in k])
if accs.size == 0 or any([acc is None for acc in accs]):
continue
mean_acc = np.mean(accs) * 100.0
eval_results.append(
EvalResult(
eval_name=result_key,
org=org,
model=model,
revision=model_sha,
results={benchmark: mean_acc},
precision=precision, # todo model_type=, weight_type=
date=config.get("submission_date")
)
)
return result_key, eval_results
def get_eval_results() -> List[EvalResult]:
json_filepaths = []
for root, dir, files in os.walk("eval-results"):
# We should only have json files in model results
if len(files) == 0 or any([not f.endswith(".json") for f in files]):
continue
# Sort the files by date
# store results by precision maybe?
try:
files.sort(key=lambda x: x.removesuffix(".json").removeprefix("results_")[:-7])
except dateutil.parser._parser.ParserError:
files = [files[-1]]
# up_to_date = files[-1]
for file in files:
json_filepaths.append(os.path.join(root, file))
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() -> List[Dict]:
eval_results = get_eval_results()
return [e.to_dict() for e in eval_results]
|