Upload curriculum_benchmark.py
Browse files- curriculum_benchmark.py +226 -0
curriculum_benchmark.py
ADDED
@@ -0,0 +1,226 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Lint as: python3
|
2 |
+
"""CURRICULUM Benchmark"""
|
3 |
+
|
4 |
+
import json
|
5 |
+
import os
|
6 |
+
|
7 |
+
import datasets
|
8 |
+
|
9 |
+
|
10 |
+
logger = datasets.logging.get_logger(__name__)
|
11 |
+
|
12 |
+
|
13 |
+
_CITATION = """\
|
14 |
+
@misc{https://doi.org/10.48550/arxiv.2204.06283,
|
15 |
+
doi = {10.48550/ARXIV.2204.06283},
|
16 |
+
url = {https://arxiv.org/abs/2204.06283},
|
17 |
+
author = {Chen, Zeming and Gao, Qiyue},
|
18 |
+
keywords = {Computation and Language (cs.CL), Artificial Intelligence (cs.AI), FOS: Computer and information sciences, FOS: Computer and information sciences},
|
19 |
+
title = {Curriculum: A Broad-Coverage Benchmark for Linguistic Phenomena in Natural Language Understanding},
|
20 |
+
publisher = {arXiv},
|
21 |
+
year = {2022},
|
22 |
+
copyright = {Creative Commons Attribution 4.0 International}
|
23 |
+
}
|
24 |
+
"""
|
25 |
+
|
26 |
+
_DESCRIPTION = """\
|
27 |
+
We introduce Curriculum as a new format of NLI benchmark for evaluation of broad-coverage linguistic phenomena.
|
28 |
+
Curriculum contains a collection of datasets that covers 36 types of major linguistic phenomena and an evaluation procedure
|
29 |
+
for diagnosing how well a language model captures reasoning skills for distinct types of linguistic phenomena.
|
30 |
+
We show that this linguistic-phenomena-driven benchmark can serve as an effective tool for diagnosing
|
31 |
+
model behavior and verifying model learning quality.
|
32 |
+
"""
|
33 |
+
|
34 |
+
_HOMEPAGE = "https://github.com/eric11eca/curriculum-ling"
|
35 |
+
_LICENSE = "CC BY-SA 3.0"
|
36 |
+
_URL = "https://github.com/eric11eca/curriculum-ling/blob/main/benchmark/tasks/"
|
37 |
+
|
38 |
+
|
39 |
+
_DESCRIPTION_MAP = {
|
40 |
+
"analytic": "analytical thinking.",
|
41 |
+
"atomic": "reasoning on commonsense knowledge graph.",
|
42 |
+
}
|
43 |
+
|
44 |
+
_TAKS_NAMES = ["analytic", "defeasible", "boolean", "comparative",
|
45 |
+
"conditional", "context_align", "control", "coreference",
|
46 |
+
"cosmoqa", "counterfactual", "counting", "drop",
|
47 |
+
"entailment_tree", "ester", "hellaswag", "hypernymy",
|
48 |
+
"hyponymy", "kg_relations", "lexical", "logiqa",
|
49 |
+
"monotonicity_infer", "negation", "ner", "physicalqa",
|
50 |
+
"puns", "quantifier", "sentiment", "socialqa",
|
51 |
+
"spatial", "sprl", "syntactic_alternation", "syntactic_variation",
|
52 |
+
"temporal", "transitive", "verbcorner", "verbnet"]
|
53 |
+
|
54 |
+
task_label_dict = {
|
55 |
+
"lexical": ["entailed", "not-entailed"],
|
56 |
+
"transitive": ["entailed", "not-entailed"],
|
57 |
+
"hypernymy": ["entailed", "not-entailed"],
|
58 |
+
"hyponymy": ["entailed", "not-entailed"],
|
59 |
+
"ner": ["entailed", "not-entailed"],
|
60 |
+
"verbnet": ["entailed", "not-entailed"],
|
61 |
+
"verbcorner": ["entailed", "not-entailed"],
|
62 |
+
"syntactic_alternation": ["entailed", "not-entailed"],
|
63 |
+
"syntactic_variation": ["entailed", "not-entailed"],
|
64 |
+
"boolean": ["entailment", "contradiction", "neutral"],
|
65 |
+
"comparative": ["entailment", "contradiction", "neutral"],
|
66 |
+
"conditional": ["entailment", "contradiction", "neutral"],
|
67 |
+
"counting": ["entailment", "contradiction", "neutral"],
|
68 |
+
"negation": ["entailment", "contradiction", "neutral"],
|
69 |
+
"quantifier": ["entailment", "contradiction", "neutral"],
|
70 |
+
"monotonicity_infer": ["entailed", "not-entailed"],
|
71 |
+
"sentiment": ["entailed", "not-entailed"],
|
72 |
+
"kg_relations": ["entailed", "not-entailed"],
|
73 |
+
"puns": ["entailed", "not-entailed"],
|
74 |
+
"coreference": ["entailed", "not-entailed"],
|
75 |
+
"context_align": ["entailed", "not-entailed"],
|
76 |
+
"sprl": ["entailed", "not-entailed"],
|
77 |
+
"analytic": ["entailed", "not-entailed"],
|
78 |
+
"entailment_tree": ["entailed", "not-entailed"],
|
79 |
+
"socialqa": ["entailed", "not-entailed"],
|
80 |
+
"physicalqa": ["entailed", "not-entailed"],
|
81 |
+
"hellaswag": ["entailed", "not-entailed"],
|
82 |
+
"cosmoqa": ["entailed", "not-entailed"],
|
83 |
+
"logiqa": ["entailed", "not-entailed"],
|
84 |
+
"ester": ["entailed", "not-entailed"],
|
85 |
+
"drop": ["entailed", "not-entailed"],
|
86 |
+
"control": ["entailment", "contradiction", "neutral"],
|
87 |
+
"spatial": ["entailed", "not-entailed"],
|
88 |
+
"temporal": ["entailed", "not-entailed"],
|
89 |
+
"defeasible": ["entailed", "not-entailed"],
|
90 |
+
"counterfactual": ["entailed", "not-entailed"]
|
91 |
+
}
|
92 |
+
|
93 |
+
|
94 |
+
def read_file(path, mode="r", **kwargs):
|
95 |
+
with open(path, mode=mode, **kwargs) as f:
|
96 |
+
return f.read()
|
97 |
+
|
98 |
+
|
99 |
+
def write_file(data, path, mode="w", **kwargs):
|
100 |
+
with open(path, mode=mode, **kwargs) as f:
|
101 |
+
f.write(data)
|
102 |
+
|
103 |
+
|
104 |
+
def read_json(path, mode="r", **kwargs):
|
105 |
+
return json.loads(read_file(path, mode=mode, **kwargs))
|
106 |
+
|
107 |
+
|
108 |
+
def write_json(data, path):
|
109 |
+
return write_file(json.dumps(data, indent=2), path)
|
110 |
+
|
111 |
+
|
112 |
+
def read_jsonl(path, mode="r", **kwargs):
|
113 |
+
# Manually open because .splitlines is different from iterating over lines
|
114 |
+
ls = []
|
115 |
+
with open(path, mode, **kwargs) as f:
|
116 |
+
for line in f:
|
117 |
+
ls.append(json.loads(line))
|
118 |
+
return ls
|
119 |
+
|
120 |
+
|
121 |
+
def write_jsonl(data, path):
|
122 |
+
assert isinstance(data, list)
|
123 |
+
lines = [to_jsonl(elem) for elem in data]
|
124 |
+
write_file("\n".join(lines), path)
|
125 |
+
|
126 |
+
|
127 |
+
def to_jsonl(data):
|
128 |
+
return json.dumps(data).replace("\n", "")
|
129 |
+
|
130 |
+
|
131 |
+
class CurriculumConfig(datasets.BuilderConfig):
|
132 |
+
"""BuilderConfig for Curriculum."""
|
133 |
+
|
134 |
+
def __init__(self, features, data_url, citation, url, label_classes=["entailed", "not-entailed"], **kwargs):
|
135 |
+
"""BuilderConfig for Curriculum.
|
136 |
+
Args:
|
137 |
+
features: `list[string]`, list of the features that will appear in the
|
138 |
+
feature dict. Should not include "label".
|
139 |
+
data_url: `string`, url to download the zip file from.
|
140 |
+
citation: `string`, citation for the data set.
|
141 |
+
url: `string`, url for information about the data set.
|
142 |
+
label_classes: `list[string]`, the list of classes for the label if the
|
143 |
+
label is present as a string. Non-string labels will be cast to either
|
144 |
+
'False' or 'True'.
|
145 |
+
**kwargs: keyword arguments forwarded to super.
|
146 |
+
"""
|
147 |
+
# Version history:
|
148 |
+
# 1.0.0: Initial version.
|
149 |
+
super(CurriculumConfig, self).__init__(
|
150 |
+
version=datasets.Version("1.0.0"), **kwargs)
|
151 |
+
self.features = features
|
152 |
+
self.label_classes = label_classes
|
153 |
+
self.data_url = data_url
|
154 |
+
self.citation = citation
|
155 |
+
self.url = url
|
156 |
+
|
157 |
+
|
158 |
+
class CurriculumBenchmark(datasets.GeneratorBasedBuilder):
|
159 |
+
"""Curriculum Benchmark. Version 1.0.0"""
|
160 |
+
|
161 |
+
BUILDER_CONFIGS = [
|
162 |
+
CurriculumConfig(
|
163 |
+
name=task_name,
|
164 |
+
description=_DESCRIPTION,
|
165 |
+
label_classes=task_label_dict[task_name],
|
166 |
+
features=["premise", "hypothesis", "idx", "gold_label"],
|
167 |
+
data_url=f"https://github.com/eric11eca/curriculum-ling/raw/main/benchmark/tasks/{task_name}.zip",
|
168 |
+
citation=_CITATION,
|
169 |
+
url="https://github.com/eric11eca/curriculum-ling/",
|
170 |
+
) for task_name in _TAKS_NAMES
|
171 |
+
]
|
172 |
+
|
173 |
+
def _info(self):
|
174 |
+
features = {feature: datasets.Value(
|
175 |
+
"string") for feature in self.config.features}
|
176 |
+
return datasets.DatasetInfo(
|
177 |
+
description=_DESCRIPTION,
|
178 |
+
features=datasets.Features(features),
|
179 |
+
supervised_keys=None,
|
180 |
+
homepage=_HOMEPAGE,
|
181 |
+
citation=_CITATION,
|
182 |
+
)
|
183 |
+
|
184 |
+
@staticmethod
|
185 |
+
def _get_filepath(dl_dir, split):
|
186 |
+
return os.path.join(dl_dir, split + ".jsonl")
|
187 |
+
|
188 |
+
def _split_generators(self, dl_manager):
|
189 |
+
dl_dir = dl_manager.download_and_extract(self.config.data_url) or ""
|
190 |
+
task_name = _get_task_name_from_data_url(self.config.data_url)
|
191 |
+
dl_dir = os.path.join(dl_dir, task_name)
|
192 |
+
|
193 |
+
return [
|
194 |
+
datasets.SplitGenerator(
|
195 |
+
name=datasets.Split.TRAIN,
|
196 |
+
gen_kwargs={
|
197 |
+
"data_file": os.path.join(dl_dir, "train.jsonl"),
|
198 |
+
"split": datasets.Split.TRAIN,
|
199 |
+
},
|
200 |
+
),
|
201 |
+
datasets.SplitGenerator(
|
202 |
+
name=datasets.Split.VALIDATION,
|
203 |
+
gen_kwargs={
|
204 |
+
"data_file": os.path.join(dl_dir, "val.jsonl"),
|
205 |
+
"split": datasets.Split.VALIDATION,
|
206 |
+
},
|
207 |
+
)
|
208 |
+
]
|
209 |
+
|
210 |
+
def _generate_examples(self, data_file, split):
|
211 |
+
"""This function returns the examples in the raw (text) form."""
|
212 |
+
logger.info("generating examples from = %s", data_file)
|
213 |
+
|
214 |
+
dataset = read_jsonl(data_file)
|
215 |
+
for id_, data in enumerate(dataset):
|
216 |
+
|
217 |
+
yield id_, {
|
218 |
+
"premise": data["premise"],
|
219 |
+
"hypothesis": data["hypothesis"],
|
220 |
+
"gold_label": data["gold_label"],
|
221 |
+
"idx": id_
|
222 |
+
}
|
223 |
+
|
224 |
+
|
225 |
+
def _get_task_name_from_data_url(data_url):
|
226 |
+
return data_url.split("/")[-1].split(".")[0]
|