Datasets:
Upload seaeval.py with huggingface_hub
Browse files- seaeval.py +238 -0
seaeval.py
ADDED
@@ -0,0 +1,238 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2022 The HuggingFace Datasets Authors and the current dataset script contributor.
|
3 |
+
#
|
4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
+
# you may not use this file except in compliance with the License.
|
6 |
+
# You may obtain a copy of the License at
|
7 |
+
#
|
8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
+
#
|
10 |
+
# Unless required by applicable law or agreed to in writing, software
|
11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
+
# See the License for the specific language governing permissions and
|
14 |
+
# limitations under the License.
|
15 |
+
|
16 |
+
from pathlib import Path
|
17 |
+
from typing import Dict, List, Tuple
|
18 |
+
|
19 |
+
import datasets
|
20 |
+
import pandas as pd
|
21 |
+
|
22 |
+
from seacrowd.utils import schemas
|
23 |
+
from seacrowd.utils.configs import SEACrowdConfig
|
24 |
+
from seacrowd.utils.constants import Licenses, Tasks
|
25 |
+
|
26 |
+
_CITATION = """\
|
27 |
+
@article{SeaEval2023,
|
28 |
+
title={SeaEval for Multilingual Foundation Models: From Cross-Lingual Alignment to Cultural Reasoning},
|
29 |
+
author={Wang, Bin and Liu, Zhengyuan and Huang, Xin and Jiao, Fangkai and Ding, Yang and Aw, Ai Ti and Chen, Nancy F.},
|
30 |
+
journal={arXiv preprint arXiv:2309.04766},
|
31 |
+
year={2023},
|
32 |
+
url={https://github.com/SeaEval/SeaEval}
|
33 |
+
}
|
34 |
+
"""
|
35 |
+
|
36 |
+
_DATASETNAME = "seaeval"
|
37 |
+
|
38 |
+
_DESCRIPTION = """\
|
39 |
+
SeaEval is a benchmark toolkit for evaluating multilingual LLMs. The benchmark contains 28 datasets,
|
40 |
+
covering 7 languages. It contains 2 datasets for cross-lingual consistency, each containing parallel
|
41 |
+
questions for the 7 represented languages. It alsoc ontains 4 datasets for cultural reasoning
|
42 |
+
(multiple choice Q&A) that are in English but focused on regions including Singapore and Philipines.
|
43 |
+
|
44 |
+
This dataloader provides examples for Indonesia, Vietnamese, Malay, and Filipino.
|
45 |
+
"""
|
46 |
+
|
47 |
+
_HOMEPAGE = "https://github.com/SeaEval/SeaEval"
|
48 |
+
|
49 |
+
_LANGUAGES = {"ind": "Indonesian", "vie": "Vietnamese", "zlm": "Malay", "fil": "Filipino"}
|
50 |
+
|
51 |
+
_LICENSE = Licenses.CC_BY_NC_4_0.value
|
52 |
+
|
53 |
+
_LOCAL = False
|
54 |
+
|
55 |
+
_URLS = {
|
56 |
+
"cross_mmlu": "https://huggingface.co/datasets/SeaEval/SeaEval_datasets/raw/main/cross_mmlu.json",
|
57 |
+
"cross_logiqa": "https://huggingface.co/datasets/SeaEval/SeaEval_datasets/raw/main/cross_logiqa.json",
|
58 |
+
"sg_eval": "https://huggingface.co/datasets/SeaEval/SeaEval_datasets/raw/main/sg_eval.json",
|
59 |
+
"ph_eval": "https://huggingface.co/datasets/SeaEval/SeaEval_datasets/raw/main/ph_eval.json",
|
60 |
+
}
|
61 |
+
|
62 |
+
_SUPPORTED_TASKS = [Tasks.COMMONSENSE_REASONING, Tasks.QUESTION_ANSWERING]
|
63 |
+
|
64 |
+
_SOURCE_VERSION = "1.0.0"
|
65 |
+
|
66 |
+
_SEACROWD_VERSION = "2024.06.20"
|
67 |
+
|
68 |
+
|
69 |
+
class SeaEvalDataset(datasets.GeneratorBasedBuilder):
|
70 |
+
"""
|
71 |
+
SeaEval is a benchmark for evaluating multilingual LLMs from https://github.com/SeaEval/SeaEval.
|
72 |
+
"""
|
73 |
+
|
74 |
+
SOURCE_VERSION = datasets.Version(_SOURCE_VERSION)
|
75 |
+
SEACROWD_VERSION = datasets.Version(_SEACROWD_VERSION)
|
76 |
+
|
77 |
+
LANGUAGES_EXCHANGED = dict((v, k) for k, v in _LANGUAGES.items())
|
78 |
+
SUBSETS_CROSS_MMLU = ["cross_mmlu_" + lang for lang in _LANGUAGES.keys()]
|
79 |
+
SUBSETS_CROSS_LOGIQA = ["cross_logiqa_" + lang for lang in _LANGUAGES.keys()]
|
80 |
+
SUBSETS = SUBSETS_CROSS_MMLU + SUBSETS_CROSS_LOGIQA + ["sg_eval_eng", "ph_eval_eng"]
|
81 |
+
|
82 |
+
BUILDER_CONFIGS = [
|
83 |
+
SEACrowdConfig(
|
84 |
+
name=f"{_DATASETNAME}_{subset}_source",
|
85 |
+
version=datasets.Version(_SOURCE_VERSION),
|
86 |
+
description=f"{_DATASETNAME}_{subset} source schema",
|
87 |
+
schema="source",
|
88 |
+
subset_id=f"{_DATASETNAME}_{subset}",
|
89 |
+
)
|
90 |
+
for subset in SUBSETS
|
91 |
+
]
|
92 |
+
|
93 |
+
BUILDER_CONFIGS += [
|
94 |
+
SEACrowdConfig(
|
95 |
+
name=f"{_DATASETNAME}_{subset}_seacrowd_qa",
|
96 |
+
version=datasets.Version(_SOURCE_VERSION),
|
97 |
+
description=f"{_DATASETNAME}_{subset} SEACrowd schema",
|
98 |
+
schema="seacrowd_qa",
|
99 |
+
subset_id=f"{_DATASETNAME}_{subset}",
|
100 |
+
)
|
101 |
+
for subset in SUBSETS
|
102 |
+
]
|
103 |
+
|
104 |
+
def _info(self) -> datasets.DatasetInfo:
|
105 |
+
if self.config.schema == "source" and self.config.subset_id not in ["cross_logiqa", "ph_eval"]:
|
106 |
+
features = datasets.Features(
|
107 |
+
{
|
108 |
+
"id": datasets.Value("string"),
|
109 |
+
"question": datasets.Value("string"),
|
110 |
+
"choices": datasets.Sequence(datasets.Value("string")),
|
111 |
+
"answer": datasets.Value("string"),
|
112 |
+
}
|
113 |
+
)
|
114 |
+
elif self.config.schema == "source" and self.config.subset_id == "cross_logiqa":
|
115 |
+
features = datasets.Features(
|
116 |
+
{
|
117 |
+
"id": datasets.Value("string"),
|
118 |
+
"question": datasets.Value("string"),
|
119 |
+
"context": datasets.Value("string"),
|
120 |
+
"choices": datasets.Sequence(datasets.Value("string")),
|
121 |
+
"answer": datasets.Value("string"),
|
122 |
+
}
|
123 |
+
)
|
124 |
+
elif self.config.schema == "source" and self.config.subset_id == "ph_eval":
|
125 |
+
features = datasets.Features(
|
126 |
+
{
|
127 |
+
"id": datasets.Value("string"),
|
128 |
+
"question": datasets.Value("string"),
|
129 |
+
"choices": datasets.Sequence(datasets.Value("string")),
|
130 |
+
"answer": datasets.Value("string"),
|
131 |
+
"category": datasets.Value("string"),
|
132 |
+
}
|
133 |
+
)
|
134 |
+
elif self.config.schema == "seacrowd_qa":
|
135 |
+
features = schemas.qa_features
|
136 |
+
else:
|
137 |
+
raise ValueError(f"Unexpected schema received! {self.config.schema}")
|
138 |
+
|
139 |
+
return datasets.DatasetInfo(
|
140 |
+
description=_DESCRIPTION,
|
141 |
+
features=features,
|
142 |
+
homepage=_HOMEPAGE,
|
143 |
+
license=_LICENSE,
|
144 |
+
citation=_CITATION,
|
145 |
+
)
|
146 |
+
|
147 |
+
def _split_generators(self, dl_manager: datasets.DownloadManager) -> List[datasets.SplitGenerator]:
|
148 |
+
"""
|
149 |
+
Returns SplitGenerators.
|
150 |
+
"""
|
151 |
+
|
152 |
+
data = {key: dl_manager.download_and_extract(value) for key, value in _URLS.items()}
|
153 |
+
|
154 |
+
paths = {}
|
155 |
+
file = self.config.subset_id.split("_")
|
156 |
+
file = "_".join(file[1:3])
|
157 |
+
paths[self.config.subset_id] = data[file]
|
158 |
+
|
159 |
+
return [
|
160 |
+
datasets.SplitGenerator(
|
161 |
+
name=datasets.Split.TEST,
|
162 |
+
gen_kwargs={
|
163 |
+
"paths": paths,
|
164 |
+
"split": "test",
|
165 |
+
},
|
166 |
+
),
|
167 |
+
]
|
168 |
+
|
169 |
+
def _generate_examples(self, paths: Path, split: str) -> Tuple[int, Dict]:
|
170 |
+
"""
|
171 |
+
Yields examples as (key, example) tuples.
|
172 |
+
"""
|
173 |
+
|
174 |
+
language = self.config.subset_id.split("_")[3]
|
175 |
+
examples = None
|
176 |
+
|
177 |
+
for key, path in paths.items():
|
178 |
+
if "cross" in key:
|
179 |
+
data = pd.read_json(path).rename(columns=self.LANGUAGES_EXCHANGED)
|
180 |
+
data = pd.melt(data, id_vars=["id"], value_vars=_LANGUAGES.keys(), var_name="language")
|
181 |
+
data_flattened = pd.json_normalize(data["value"])
|
182 |
+
data_merged = pd.merge(data, data_flattened, left_index=True, right_index=True)
|
183 |
+
data_filtered = data_merged[data_merged["language"] == language].drop(columns=["value", "language"])
|
184 |
+
examples = data_filtered.to_records()
|
185 |
+
elif "eval" in key:
|
186 |
+
data = pd.read_json(path)
|
187 |
+
examples = data.to_records()
|
188 |
+
|
189 |
+
idx = 0
|
190 |
+
if self.config.schema == "source" and self.config.subset_id not in ["cross_logiqa", "ph_eval"]:
|
191 |
+
for row in examples:
|
192 |
+
x = {
|
193 |
+
"id": row["id"],
|
194 |
+
"question": row["question"],
|
195 |
+
"choices": row["choices"],
|
196 |
+
"answer": row["answer"],
|
197 |
+
}
|
198 |
+
yield idx, x
|
199 |
+
idx += 1
|
200 |
+
elif self.config.schema == "source" and self.config.subset_id == "cross_logiqa":
|
201 |
+
for row in examples:
|
202 |
+
x = {
|
203 |
+
"id": row["id"],
|
204 |
+
"question": row["question"],
|
205 |
+
"context": row["context"] if "context" in row else None,
|
206 |
+
"choices": row["choices"],
|
207 |
+
"answer": row["answer"],
|
208 |
+
}
|
209 |
+
yield idx, x
|
210 |
+
idx += 1
|
211 |
+
elif self.config.schema == "source" and self.config.subset_id == "ph_eval":
|
212 |
+
for row in examples:
|
213 |
+
x = {
|
214 |
+
"id": row["id"],
|
215 |
+
"question": row["question"],
|
216 |
+
"choices": row["choices"],
|
217 |
+
"answer": row["answer"],
|
218 |
+
"category": row["category"] if "category" in row else None,
|
219 |
+
}
|
220 |
+
yield idx, x
|
221 |
+
idx += 1
|
222 |
+
elif self.config.schema == "seacrowd_qa":
|
223 |
+
for row in examples:
|
224 |
+
x = {
|
225 |
+
"id": idx,
|
226 |
+
"question_id": row["id"],
|
227 |
+
"document_id": row["id"],
|
228 |
+
"question": row["question"],
|
229 |
+
"type": "multiple_choice",
|
230 |
+
"choices": row["choices"],
|
231 |
+
"context": row["context"] if "context" in row else None,
|
232 |
+
"answer": [row["answer"]],
|
233 |
+
"meta": {},
|
234 |
+
}
|
235 |
+
yield idx, x
|
236 |
+
idx += 1
|
237 |
+
else:
|
238 |
+
raise ValueError(f"Invalid schema: {self.config.schema}")
|