forked from openai/simple-evals
-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathmgsm_eval.py
196 lines (159 loc) · 9.79 KB
/
mgsm_eval.py
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
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
"""
MGSM: Multilingual Grade School Math Benchmark (MGSM) is a benchmark of grade-school math problems.
Language Models are Multilingual Chain-of-Thought Reasoners
Freda Shi, Mirac Suzgun, Markus Freitag, Xuezhi Wang, Suraj Srivats, Soroush Vosoughi, Hyung Won Chung, Yi Tay, Sebastian Ruder, Denny Zhou, Dipanjan Das, Jason Wei
https://arxiv.org/abs/2210.03057 reference: https://github.com/google-research/url-nlp
"""
import re
from typing import Optional
from data_utils import download_file
from mmlu_eval import HTML_JINJA
from eval_types import Eval, EvalResult, SamplerBase, SingleEvalResult
from common import (
HTML_JINJA,
jinja_env,
map_with_progress,
aggregate_results,
)
ALL_LANGUAGES = ["bn", "de", "en", "es", "fr", "ja", "ru", "sw", "te", "th", "zh"]
LATIN_LANGUAGES = ["de", "en", "es", "fr", "sw"]
NON_LATIN_LANGUAGES = ["bn", "ja", "ru", "te", "th", "zh"]
LANG_TO_FPATH = {
"bn": "https://openaipublic.blob.core.windows.net/simple-evals/mgsm_bn.tsv",
"de": "https://openaipublic.blob.core.windows.net/simple-evals/mgsm_de.tsv",
"en": "https://openaipublic.blob.core.windows.net/simple-evals/mgsm_en.tsv",
"es": "https://openaipublic.blob.core.windows.net/simple-evals/mgsm_es.tsv",
"fr": "https://openaipublic.blob.core.windows.net/simple-evals/mgsm_fr.tsv",
"ja": "https://openaipublic.blob.core.windows.net/simple-evals/mgsm_ja.tsv",
"ru": "https://openaipublic.blob.core.windows.net/simple-evals/mgsm_ru.tsv",
"sw": "https://openaipublic.blob.core.windows.net/simple-evals/mgsm_sw.tsv",
"te": "https://openaipublic.blob.core.windows.net/simple-evals/mgsm_te.tsv",
"th": "https://openaipublic.blob.core.windows.net/simple-evals/mgsm_th.tsv",
"zh": "https://openaipublic.blob.core.windows.net/simple-evals/mgsm_zh.tsv",
}
LANG_TO_INSTRUCTIONS = {
"en": """Solve this math problem. Give the reasoning steps before giving the final answer on the last line by itself in the format of "Answer:". Do not add anything other than the integer answer after "Answer:".
{input}""",
"bn": """এই গণিতের সমস্যাটি সমাধান করুন। চূড়ান্ত উত্তর দেওয়ার আগে যুক্তিসম্পন্ন পদক্ষেপ প্রদান করুন। চূড়ান্ত উত্তরটি একক সংখ্যা হিসাবে "উত্তর:" এর পরে শেষ লাইনে দিন। "উত্তর:" এর পরে অন্য কিছু যুক্ত করবেন না।.
{input}""",
"de": """Löse dieses Mathematikproblem. Gib die Schritte zur Begründung an, bevor du die endgültige Antwort in der letzten Zeile alleine im Format "Antwort:" gibst. Füge nichts anderes als die ganzzahlige Antwort nach "Antwort:" hinzu.
{input}""",
"es": """Resuelve este problema matemático. Proporciona los pasos de razonamiento antes de dar la respuesta final en la última línea por sí misma en el formato de "Respuesta:". No añadas nada más que la respuesta entera después de "Respuesta:".
{input}""",
"fr": """Résolvez ce problème de mathématiques. Donnez les étapes de raisonnement avant de fournir la réponse finale sur la dernière ligne elle-même dans le format de "Réponse:". N'ajoutez rien d'autre que la réponse entière après "Réponse:".
{input}""",
"ja": """の数学の問題を解いてください。最終的な答えを出す前に、解答の推論過程を記述してください。そして最後の行には "答え:" の形式で答えを記述し、その後には整数の答え以外何も追加しないでください。
{input}""",
"ru": """Решите эту математическую задачу. Объясните шаги рассуждения перед тем, как дать окончательный ответ в последней строке сам по себе в формате "Ответ:". Не добавляйте ничего, кроме целочисленного ответа после "Ответ:".
{input}""",
"sw": """Suluhisha tatizo hili la hesabu. Toa hatua za mantiki kabla ya kutoa jibu la mwisho kwenye mstari wa mwisho peke yake katika muundo wa "Jibu:". Usiongeze chochote kingine isipokuwa jibu la integer baada ya "Jibu:".
{input}""",
"te": """ఈ గణిత సమస్యను పరిష్కరించండి. చివరి సమాధానాన్ని ఇవ్వదానికి ముందు తర్కాత్మక అదుగులను ఇవ్వండి. చివరి పంక్తిలో మాత్రమే 'సమాధానం:' అనే ఆకారంలో చివరి సమాధానాద్ని ఇవ్వండి సమాధానం: తర్వాత పూర్ణాంక సమాధానానికి తప్పించి ఎదేనా చేర్చవద్దు.
{input}""",
"th": """แก้ปัญหาคณิตศาสตร์นี้ ให้ให้ขั้นตอนการใช้เหตุผลก่อนที่จะให้คำตอบสุดท้ายในบรรทัดสุดท้ายโดยอยู่ในรูปแบบ "คำตอบ:" ไม่ควรเพิ่มอะไรนอกจากคำตอบที่เป็นจำนวนเต็มหลังจาก "คำตอบ:"
{input}""",
"zh": """解决这个数学问题。在最后一行给出答案前，请提供推理步骤。最后一行应该以 "答案: " 的形式独立给出答案。在 "答案：" 后不要添加除整数答案之外的任何内容。
{input}""",
}
LANG_TO_ANSWER_PREFIX = {
"en": "Answer",
"bn": "উত্তর",
"de": "Antwort",
"es": "Respuesta",
"fr": "Réponse",
"ja": "答え",
"ru": "Ответ",
"sw": "Jibu",
"te": "సమాధానం",
"th": "คำตอบ",
"zh": "答案",
}
def parse_answer(answer: str, answer_prefix: str) -> str:
if answer_prefix not in answer:
return ""
answer_text = answer.split(answer_prefix)[-1].strip()
# find all the numbers (including decimals) in the string
numbers = re.findall(r"\d+\.?\d*", answer_text.replace(",", ""))
# return the first number (removing trailing decimal point if present),
# or an empty string if there were no numbers
return numbers[-1].rstrip(".") if numbers else ""
def score_mgsm(target: str, prediction: str) -> bool:
if "." in prediction:
prediction = prediction.rstrip("0").rstrip(".")
target = target.replace(",", "")
prediction = prediction.replace(",", "")
return target == prediction
def get_lang_examples(lang: str) -> list[dict[str, str]]:
fpath = LANG_TO_FPATH[lang]
examples = []
path = download_file(fpath)
with open(path, "r", encoding="utf-8") as f:
for line in f:
inputs, targets = line.strip().split("\t")
if "." in targets:
raise ValueError(f"targets {targets} contains a decimal point.")
examples.append({"inputs": inputs, "targets": targets, "lang": lang})
return examples
def get_all_examples() -> list[dict[str, str]]:
examples = []
for lang in ALL_LANGUAGES:
if lang != "en":
continue
examples += get_lang_examples(lang)
return examples
class MGSMEval(Eval):
def __init__(
self,
num_examples_per_lang: int = 250, # restrict to a subset of the data for debugging
languages: Optional[list[str]] = ALL_LANGUAGES,
):
if languages is None:
languages = ALL_LANGUAGES
else:
for language in languages:
if language not in ALL_LANGUAGES:
raise ValueError(
f"language {language} is not a valid language. "
f"It should be one in {ALL_LANGUAGES}"
)
self._languages = languages
self._num_examples_per_lang = num_examples_per_lang
examples = []
for lang in self._languages:
lang_examples = get_lang_examples(lang)
examples.extend(lang_examples[: self._num_examples_per_lang])
self.examples = examples
def __call__(self, sampler: SamplerBase) -> EvalResult:
def fn(example: dict[str, str]):
language = example["lang"]
latin_language = "group_latin" if language in LATIN_LANGUAGES else "group_non_latin"
correct_answer = example["targets"]
instructoin = LANG_TO_INSTRUCTIONS[language]
prompt_messages = [
sampler._pack_message(
content=instructoin.format(input=example["inputs"]), role="user"
)
]
try:
response_text = sampler(prompt_messages)
except Exception as e:
response_text = ""
answer_prefix = LANG_TO_ANSWER_PREFIX[language]
extracted_answer = parse_answer(response_text, answer_prefix)
score = score_mgsm(correct_answer, extracted_answer)
html = jinja_env.from_string(HTML_JINJA).render(
prompt_messages=prompt_messages,
next_message=dict(content=response_text, role="assistant"),
score=score,
correct_answer=correct_answer,
extracted_answer=extracted_answer,
)
convo = prompt_messages + [dict(content=response_text, role="assistant")]
return SingleEvalResult(
html=html,
score=score,
convo=convo,
metrics={language: score, latin_language: score},
)
results = map_with_progress(fn, self.examples)
return aggregate_results(results, default_stats=("mean", "std"))