Datasets:
ArXiv:
License:
Upload m3exam.py with huggingface_hub
Browse files
m3exam.py
ADDED
@@ -0,0 +1,318 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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 |
+
import json
|
17 |
+
import os
|
18 |
+
import re
|
19 |
+
import zipfile
|
20 |
+
from pathlib import Path
|
21 |
+
from typing import Dict, List, Tuple
|
22 |
+
|
23 |
+
import datasets
|
24 |
+
|
25 |
+
from seacrowd.utils import schemas
|
26 |
+
from seacrowd.utils.configs import SEACrowdConfig
|
27 |
+
from seacrowd.utils.constants import Licenses, Tasks
|
28 |
+
|
29 |
+
_CITATION = """\
|
30 |
+
@article{zhang2023m3exam,
|
31 |
+
title={M3Exam: A Multilingual, Multimodal, Multilevel Benchmark for Examining Large Language Models},
|
32 |
+
author={Wenxuan Zhang and Sharifah Mahani Aljunied and Chang Gao and Yew Ken Chia and Lidong Bing},
|
33 |
+
year={2023},
|
34 |
+
eprint={2306.05179},
|
35 |
+
archivePrefix={arXiv},
|
36 |
+
primaryClass={cs.CL}
|
37 |
+
}
|
38 |
+
"""
|
39 |
+
|
40 |
+
_DATASETNAME = "m3exam"
|
41 |
+
|
42 |
+
_DESCRIPTION = """\
|
43 |
+
M3Exam is a novel benchmark sourced from real and official human exam questions for evaluating LLMs\
|
44 |
+
in a multilingual, multimodal, and multilevel context. In total, M3Exam contains 12,317 questions in 9\
|
45 |
+
diverse languages with three educational levels, where about 23% of the questions require processing images\
|
46 |
+
for successful solving. M3Exam dataset covers 3 languages spoken in Southeast Asia.
|
47 |
+
"""
|
48 |
+
|
49 |
+
_HOMEPAGE = "https://github.com/DAMO-NLP-SG/M3Exam"
|
50 |
+
|
51 |
+
_LANGUAGES = ["jav", "tha", "vie"]
|
52 |
+
_LANG_MAPPER = {"jav": "javanese", "tha": "thai", "vie": "vietnamese"}
|
53 |
+
_LICENSE = Licenses.CC_BY_NC_SA_4_0.value
|
54 |
+
|
55 |
+
_LOCAL = False
|
56 |
+
_PASSWORD = "12317".encode("utf-8") # password to unzip dataset after downloading
|
57 |
+
_URLS = {
|
58 |
+
_DATASETNAME: "https://drive.usercontent.google.com/download?id=1eREETRklmXJLXrNPTyHxQ3RFdPhq_Nes&authuser=0&confirm=t",
|
59 |
+
}
|
60 |
+
|
61 |
+
_SUPPORTED_TASKS = [Tasks.QUESTION_ANSWERING, Tasks.VISUAL_QUESTION_ANSWERING]
|
62 |
+
|
63 |
+
_SOURCE_VERSION = "1.0.0"
|
64 |
+
|
65 |
+
_SEACROWD_VERSION = "2024.06.20"
|
66 |
+
|
67 |
+
|
68 |
+
class M3ExamDataset(datasets.GeneratorBasedBuilder):
|
69 |
+
"""
|
70 |
+
M3Exam is a novel benchmark sourced from real and official human exam questions for evaluating LLMs
|
71 |
+
in a multilingual, multimodal, and multilevel context. In total, M3Exam contains 12,317 questions in 9
|
72 |
+
diverse languages with three educational levels, where about 23% of the questions require processing images
|
73 |
+
for successful solving. M3Exam dataset covers 3 languages spoken in Southeast Asia.
|
74 |
+
"""
|
75 |
+
|
76 |
+
SOURCE_VERSION = datasets.Version(_SOURCE_VERSION)
|
77 |
+
SEACROWD_VERSION = datasets.Version(_SEACROWD_VERSION)
|
78 |
+
|
79 |
+
BUILDER_CONFIGS = (
|
80 |
+
[SEACrowdConfig(name=f"{_DATASETNAME}_{lang}_source", version=datasets.Version(_SOURCE_VERSION), description=f"{_DATASETNAME} source schema", schema="source", subset_id=f"{_DATASETNAME}") for lang in _LANGUAGES]
|
81 |
+
+ [
|
82 |
+
SEACrowdConfig(
|
83 |
+
name=f"{_DATASETNAME}_{lang}_seacrowd_qa",
|
84 |
+
version=datasets.Version(_SEACROWD_VERSION),
|
85 |
+
description=f"{_DATASETNAME} SEACrowd schema",
|
86 |
+
schema="seacrowd_qa",
|
87 |
+
subset_id=f"{_DATASETNAME}",
|
88 |
+
)
|
89 |
+
for lang in _LANGUAGES
|
90 |
+
]
|
91 |
+
+ [
|
92 |
+
SEACrowdConfig(
|
93 |
+
name=f"{_DATASETNAME}_{lang}_seacrowd_imqa",
|
94 |
+
version=datasets.Version(_SEACROWD_VERSION),
|
95 |
+
description=f"{_DATASETNAME} SEACrowd schema",
|
96 |
+
schema="seacrowd_imqa",
|
97 |
+
subset_id=f"{_DATASETNAME}",
|
98 |
+
)
|
99 |
+
for lang in _LANGUAGES
|
100 |
+
]
|
101 |
+
)
|
102 |
+
|
103 |
+
DEFAULT_CONFIG_NAME = f"{_DATASETNAME}_jav_source"
|
104 |
+
|
105 |
+
def _info(self) -> datasets.DatasetInfo:
|
106 |
+
|
107 |
+
if self.config.schema == "source":
|
108 |
+
features = datasets.Features(
|
109 |
+
{
|
110 |
+
"question_text": datasets.Value("string"),
|
111 |
+
"background_description": datasets.Sequence(datasets.Value("string")),
|
112 |
+
"answer_text": datasets.Value("string"),
|
113 |
+
"options": datasets.Sequence(datasets.Value("string")),
|
114 |
+
"language": datasets.Value("string"),
|
115 |
+
"level": datasets.Value("string"),
|
116 |
+
"subject": datasets.Value("string"),
|
117 |
+
"subject_category": datasets.Value("string"),
|
118 |
+
"year": datasets.Value("string"),
|
119 |
+
"need_image": datasets.Value("string"),
|
120 |
+
"image_paths": datasets.Sequence(datasets.Value("string")),
|
121 |
+
}
|
122 |
+
)
|
123 |
+
elif self.config.schema == "seacrowd_qa":
|
124 |
+
features = schemas.qa_features
|
125 |
+
features["meta"] = {
|
126 |
+
"background_description": datasets.Sequence(datasets.Value("string")),
|
127 |
+
"level": datasets.Value("string"),
|
128 |
+
"subject": datasets.Value("string"),
|
129 |
+
"subject_category": datasets.Value("string"),
|
130 |
+
"year": datasets.Value("string"),
|
131 |
+
}
|
132 |
+
elif self.config.schema == "seacrowd_imqa":
|
133 |
+
features = schemas.imqa_features
|
134 |
+
features["meta"] = {
|
135 |
+
"background_description": datasets.Sequence(datasets.Value("string")),
|
136 |
+
"level": datasets.Value("string"),
|
137 |
+
"subject": datasets.Value("string"),
|
138 |
+
"subject_category": datasets.Value("string"),
|
139 |
+
"year": datasets.Value("string"),
|
140 |
+
}
|
141 |
+
|
142 |
+
return datasets.DatasetInfo(
|
143 |
+
description=_DESCRIPTION,
|
144 |
+
features=features,
|
145 |
+
homepage=_HOMEPAGE,
|
146 |
+
license=_LICENSE,
|
147 |
+
citation=_CITATION,
|
148 |
+
)
|
149 |
+
|
150 |
+
def _split_generators(self, dl_manager: datasets.DownloadManager) -> List[datasets.SplitGenerator]:
|
151 |
+
"""Returns SplitGenerators."""
|
152 |
+
urls = _URLS[_DATASETNAME]
|
153 |
+
lang = self.config.name.split("_")[1]
|
154 |
+
|
155 |
+
data_dir = dl_manager.download(urls)
|
156 |
+
|
157 |
+
if not os.path.exists(data_dir + "_extracted"):
|
158 |
+
if not os.path.exists(data_dir + ".zip"):
|
159 |
+
os.rename(data_dir, data_dir + ".zip")
|
160 |
+
with zipfile.ZipFile(data_dir + ".zip", "r") as zip_ref:
|
161 |
+
zip_ref.extractall(data_dir + "_extracted", pwd=_PASSWORD) # unzipping with password
|
162 |
+
if not os.path.exists(data_dir):
|
163 |
+
os.rename(data_dir + ".zip", data_dir)
|
164 |
+
image_generator = [
|
165 |
+
datasets.SplitGenerator(
|
166 |
+
name=datasets.Split.TRAIN,
|
167 |
+
gen_kwargs={
|
168 |
+
"filepath": os.path.join(data_dir + "_extracted", "data/multimodal-question"),
|
169 |
+
"split": "train",
|
170 |
+
},
|
171 |
+
),
|
172 |
+
]
|
173 |
+
|
174 |
+
text_generator = [
|
175 |
+
datasets.SplitGenerator(
|
176 |
+
name=datasets.Split.TEST,
|
177 |
+
gen_kwargs={
|
178 |
+
"filepath": os.path.join(data_dir + "_extracted", f"data/text-question/{_LANG_MAPPER[lang]}-questions-test.json"),
|
179 |
+
"split": "test",
|
180 |
+
},
|
181 |
+
),
|
182 |
+
datasets.SplitGenerator(
|
183 |
+
name=datasets.Split.VALIDATION,
|
184 |
+
gen_kwargs={
|
185 |
+
"filepath": os.path.join(data_dir + "_extracted", f"data/text-question/{_LANG_MAPPER[lang]}-questions-dev.json"),
|
186 |
+
"split": "dev",
|
187 |
+
},
|
188 |
+
),
|
189 |
+
]
|
190 |
+
if "imqa" in self.config.name:
|
191 |
+
return image_generator
|
192 |
+
else:
|
193 |
+
if "source" in self.config.name:
|
194 |
+
image_generator.extend(text_generator)
|
195 |
+
return image_generator
|
196 |
+
else:
|
197 |
+
return text_generator
|
198 |
+
|
199 |
+
def _generate_examples(self, filepath: Path, split: str) -> Tuple[int, Dict]:
|
200 |
+
"""Yields examples as (key, example) tuples."""
|
201 |
+
lang = self.config.name.split("_")[1]
|
202 |
+
thai_answer_mapper = {"1": "1", "2": "2", "3": "3", "4": "4", "5": "5", "๑": "1", "๒": "2", "๓": "3", "๔": "4", "๕": "5"}
|
203 |
+
if self.config.schema == "source":
|
204 |
+
if split == "train":
|
205 |
+
filepath_json = os.path.join(filepath, f"{_LANG_MAPPER[lang]}-questions-image.json")
|
206 |
+
with open(filepath_json, "r") as file:
|
207 |
+
data = json.load(file)
|
208 |
+
idx = 0
|
209 |
+
for json_obj in data:
|
210 |
+
image_paths = []
|
211 |
+
for text in [json_obj["question_text"]] + json_obj["options"] + json_obj["background_description"]:
|
212 |
+
matches = re.findall(r"\[image-(\d+)\.(jpg|png)\]", text)
|
213 |
+
if matches:
|
214 |
+
image_path = [os.path.join(filepath, f"images-{_LANG_MAPPER[lang]}/image-{image_number[0]}.{image_number[1]}") for image_number in matches]
|
215 |
+
image_paths.extend(image_path)
|
216 |
+
example = {
|
217 |
+
"question_text": json_obj["question_text"],
|
218 |
+
"background_description": json_obj["background_description"] if "background_description" in json_obj.keys() else None,
|
219 |
+
"answer_text": json_obj["answer_text"],
|
220 |
+
"options": json_obj["options"],
|
221 |
+
"language": json_obj["language"] if "language" in json_obj.keys() else None,
|
222 |
+
"level": json_obj["level"] if "level" in json_obj.keys() else None,
|
223 |
+
"subject": json_obj["subject"] if "subject" in json_obj.keys() else None,
|
224 |
+
"subject_category": json_obj["subject_category"] if "subject_category" in json_obj.keys() else None,
|
225 |
+
"year": json_obj["year"] if "year" in json_obj.keys() else None,
|
226 |
+
"need_image": "yes",
|
227 |
+
"image_paths": image_paths,
|
228 |
+
}
|
229 |
+
yield idx, example
|
230 |
+
idx += 1
|
231 |
+
else:
|
232 |
+
with open(filepath, "r") as file:
|
233 |
+
data = json.load(file)
|
234 |
+
idx = 0
|
235 |
+
for json_obj in data:
|
236 |
+
example = {
|
237 |
+
"question_text": json_obj["question_text"],
|
238 |
+
"background_description": json_obj["background_description"] if "background_description" in json_obj.keys() else None,
|
239 |
+
"answer_text": json_obj["answer_text"],
|
240 |
+
"options": json_obj["options"],
|
241 |
+
"language": json_obj["language"] if "language" in json_obj.keys() else None,
|
242 |
+
"level": json_obj["level"] if "level" in json_obj.keys() else None,
|
243 |
+
"subject": json_obj["subject"] if "subject" in json_obj.keys() else None,
|
244 |
+
"subject_category": json_obj["subject_category"] if "subject_category" in json_obj.keys() else None,
|
245 |
+
"year": json_obj["year"] if "year" in json_obj.keys() else None,
|
246 |
+
"need_image": "no",
|
247 |
+
"image_paths": None,
|
248 |
+
}
|
249 |
+
yield idx, example
|
250 |
+
idx += 1
|
251 |
+
|
252 |
+
elif self.config.schema == "seacrowd_qa":
|
253 |
+
with open(filepath, "r") as file:
|
254 |
+
data = json.load(file)
|
255 |
+
idx = 0
|
256 |
+
|
257 |
+
for json_obj in data:
|
258 |
+
answer = [".".join(answer.split(".")[1:]).strip() for answer in json_obj["options"] if json_obj["answer_text"] == answer.split(".")[0]]
|
259 |
+
if "_tha_" in self.config.name and len(answer) == 0:
|
260 |
+
answer = [".".join(answer.split(".")[1:]).strip() for answer in json_obj["options"] if thai_answer_mapper[json_obj["answer_text"]] == thai_answer_mapper[answer.split(".")[0]]]
|
261 |
+
|
262 |
+
example = {
|
263 |
+
"id": idx,
|
264 |
+
"question_id": idx,
|
265 |
+
"document_id": idx,
|
266 |
+
"question": json_obj["question_text"],
|
267 |
+
"type": "multiple_choice",
|
268 |
+
"choices": [".".join(answer.split(".")[1:]).strip() for answer in json_obj["options"]],
|
269 |
+
"context": "",
|
270 |
+
"answer": answer,
|
271 |
+
"meta": {
|
272 |
+
"background_description": json_obj["background_description"] if "background_description" in json_obj.keys() else None,
|
273 |
+
"level": json_obj["level"] if "level" in json_obj.keys() else None,
|
274 |
+
"subject": json_obj["subject"] if "subject" in json_obj.keys() else None,
|
275 |
+
"subject_category": json_obj["subject_category"] if "subject_category" in json_obj.keys() else None,
|
276 |
+
"year": json_obj["year"] if "year" in json_obj.keys() else None,
|
277 |
+
},
|
278 |
+
}
|
279 |
+
yield idx, example
|
280 |
+
idx += 1
|
281 |
+
|
282 |
+
elif self.config.schema == "seacrowd_imqa":
|
283 |
+
filepath_json = os.path.join(filepath, f"{_LANG_MAPPER[lang]}-questions-image.json")
|
284 |
+
with open(filepath_json, "r") as file:
|
285 |
+
data = json.load(file)
|
286 |
+
idx = 0
|
287 |
+
|
288 |
+
for json_obj in data:
|
289 |
+
answer = [".".join(answer.split(".")[1:]).strip() for answer in json_obj["options"] if json_obj["answer_text"] == answer.split(".")[0]]
|
290 |
+
if "_tha_" in self.config.name and len(answer) == 0:
|
291 |
+
answer = [".".join(answer.split(".")[1:]).strip() for answer in json_obj["options"] if thai_answer_mapper[json_obj["answer_text"]] == thai_answer_mapper[answer.split(".")[0]]]
|
292 |
+
image_paths = []
|
293 |
+
for text in [json_obj["question_text"]] + json_obj["options"] + json_obj["background_description"]:
|
294 |
+
matches = re.findall(r"\[image-(\d+)\.(jpg|png)\]", text)
|
295 |
+
if matches:
|
296 |
+
image_path = [os.path.join(filepath, f"images-{_LANG_MAPPER[lang]}/image-{image_number[0]}.{image_number[1]}") for image_number in matches]
|
297 |
+
image_paths.extend(image_path)
|
298 |
+
|
299 |
+
example = {
|
300 |
+
"id": idx,
|
301 |
+
"question_id": idx,
|
302 |
+
"document_id": idx,
|
303 |
+
"questions": [json_obj["question_text"]],
|
304 |
+
"type": "multiple_choice",
|
305 |
+
"choices": [".".join(answer.split(".")[1:]).strip() for answer in json_obj["options"]],
|
306 |
+
"context": "",
|
307 |
+
"answer": answer,
|
308 |
+
"image_paths": image_paths,
|
309 |
+
"meta": {
|
310 |
+
"background_description": json_obj["background_description"] if "background_description" in json_obj.keys() else None,
|
311 |
+
"level": json_obj["level"] if "level" in json_obj.keys() else None,
|
312 |
+
"subject": json_obj["subject"] if "subject" in json_obj.keys() else None,
|
313 |
+
"subject_category": json_obj["subject_category"] if "subject_category" in json_obj.keys() else None,
|
314 |
+
"year": json_obj["year"] if "year" in json_obj.keys() else None,
|
315 |
+
},
|
316 |
+
}
|
317 |
+
yield idx, example
|
318 |
+
idx += 1
|