Upload 3 files
Browse files- evaluation/README.md +10 -0
- evaluation/calculate_scores.py +79 -0
- evaluation/grader.py +600 -0
evaluation/README.md
ADDED
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
|
3 |
+
|
4 |
+
## Introduction
|
5 |
+
This is the evaluation script used to reproduce math benchmarks scores for AceMath-1.5B/7B/72B-Instruct models based on their outputs. The benchmark can be downloaded from [Qwen2.5-Math](https://github.com/QwenLM/Qwen2.5-Math/tree/main/evaluation/data).
|
6 |
+
|
7 |
+
## Calculate Scores
|
8 |
+
```console
|
9 |
+
python calculate_scores.py
|
10 |
+
```
|
evaluation/calculate_scores.py
ADDED
@@ -0,0 +1,79 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
from grader import is_equal
|
3 |
+
import json
|
4 |
+
import re
|
5 |
+
|
6 |
+
|
7 |
+
def get_gold_list(datapath, dataset_name):
|
8 |
+
|
9 |
+
assert dataset_name in ["gsm8k", "math", "minerva_math", "gaokao2023en", "olympiadbench", "collegemath"]
|
10 |
+
|
11 |
+
gold_list = []
|
12 |
+
with open(datapath, "r") as f:
|
13 |
+
for line in f:
|
14 |
+
item = json.loads(line)
|
15 |
+
|
16 |
+
if dataset_name == "gsm8k":
|
17 |
+
gold = item['answer'].split("#### ")[-1]
|
18 |
+
|
19 |
+
elif dataset_name == "math":
|
20 |
+
gold = item['answer']
|
21 |
+
|
22 |
+
elif dataset_name == "minerva_math":
|
23 |
+
pattern = r"\\boxed\{((?:[^{}]|\{(?:[^{}]|\{[^{}]*\})*\})*)\}"
|
24 |
+
pattern_re = re.compile(pattern, re.DOTALL)
|
25 |
+
solution = item['solution']
|
26 |
+
matches = pattern_re.findall(solution)
|
27 |
+
if len(matches) == 0:
|
28 |
+
gold = None
|
29 |
+
else:
|
30 |
+
gold = matches[-1]
|
31 |
+
|
32 |
+
elif dataset_name == "gaokao2023en":
|
33 |
+
gold = re.sub(r'^\$(.*)\$$', r'\1', item['answer'])
|
34 |
+
|
35 |
+
elif dataset_name == "olympiadbench":
|
36 |
+
gold = re.sub(r'^\$(.*)\$$', r'\1', item['final_answer'][0])
|
37 |
+
|
38 |
+
elif dataset_name == "collegemath":
|
39 |
+
gold = re.sub(r'^\$(.*)\$$', r'\1', item['answer'])
|
40 |
+
|
41 |
+
gold_list.append(gold)
|
42 |
+
|
43 |
+
return gold_list
|
44 |
+
|
45 |
+
|
46 |
+
def get_scores_on_math_benchmarks(model_output_path, test_gold_path, dataset_name):
|
47 |
+
|
48 |
+
gold_list = get_gold_list(test_gold_path, dataset_name)
|
49 |
+
|
50 |
+
"""TODO
|
51 |
+
Get the output_list from model_output_path
|
52 |
+
output_list is a list of string (List[str])
|
53 |
+
Each string represents the model's response for a corresponding question in the benchmark
|
54 |
+
Therefore, the length of output_list must match the length of gold_list.
|
55 |
+
|
56 |
+
output_list = ...
|
57 |
+
"""
|
58 |
+
|
59 |
+
correct = 0
|
60 |
+
for output, gold in zip(output_list, gold_list):
|
61 |
+
if is_equal(output, gold, dataset_name):
|
62 |
+
correct += 1
|
63 |
+
|
64 |
+
print("accuracy on %s is %.4f" % (dataset_name, correct / len(gold_list)))
|
65 |
+
|
66 |
+
|
67 |
+
if __name__ == "__main__":
|
68 |
+
"""TODO
|
69 |
+
Download test benchmarks from Qwen2.5-Math
|
70 |
+
https://github.com/QwenLM/Qwen2.5-Math/tree/main/evaluation/data
|
71 |
+
|
72 |
+
Prepare model_output_path and test_gold_path for each dataset
|
73 |
+
"""
|
74 |
+
|
75 |
+
test_gold_path = "PATH_OF_THE_BENCHMARK"
|
76 |
+
model_output_path = "PATH_OF_YOUR_MODEL_OUTPUTS"
|
77 |
+
dataset_name = "DATASET_NAME" # e.g., gsm8k, math, "minerva_math", "gaokao2023en", "olympiadbench", "collegemath"
|
78 |
+
|
79 |
+
get_scores_on_math_benchmarks(model_output_path, test_gold_path, dataset_name)
|
evaluation/grader.py
ADDED
@@ -0,0 +1,600 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
"""
|
3 |
+
This script is adapted from Qwen2.5-Math
|
4 |
+
https://github.com/QwenLM/Qwen2.5-Math/blob/main/evaluation/grader.py
|
5 |
+
"""
|
6 |
+
|
7 |
+
import re
|
8 |
+
import regex
|
9 |
+
import multiprocessing
|
10 |
+
from math import isclose
|
11 |
+
from typing import Union
|
12 |
+
from collections import defaultdict
|
13 |
+
|
14 |
+
from sympy import simplify, N
|
15 |
+
from sympy.parsing.sympy_parser import parse_expr
|
16 |
+
from sympy.parsing.latex import parse_latex
|
17 |
+
|
18 |
+
|
19 |
+
def latex2sympy(sympy: str, variable_values={}):
|
20 |
+
# record frac
|
21 |
+
global frac_type
|
22 |
+
if sympy.find(r'\frac') != -1:
|
23 |
+
frac_type = r'\frac'
|
24 |
+
if sympy.find(r'\dfrac') != -1:
|
25 |
+
frac_type = r'\dfrac'
|
26 |
+
if sympy.find(r'\tfrac') != -1:
|
27 |
+
frac_type = r'\tfrac'
|
28 |
+
sympy = sympy.replace(r'\dfrac', r'\frac')
|
29 |
+
sympy = sympy.replace(r'\tfrac', r'\frac')
|
30 |
+
# Translate Transpose
|
31 |
+
sympy = sympy.replace(r'\mathrm{T}', 'T', -1)
|
32 |
+
# Translate Derivative
|
33 |
+
sympy = sympy.replace(r'\mathrm{d}', 'd', -1).replace(r'{\rm d}', 'd', -1)
|
34 |
+
# Translate Matrix
|
35 |
+
sympy = sympy.replace(r'\left[\begin{matrix}', r'\begin{bmatrix}', -1).replace(r'\end{matrix}\right]', r'\end{bmatrix}', -1)
|
36 |
+
# Translate Permutation
|
37 |
+
sympy = re.sub(r"\(([a-zA-Z0-9+\-*/\\ ]+?)\)_{([a-zA-Z0-9+\-*/\\ ]+?)}", r"\\frac{(\1)!}{((\1)-(\2))!}", sympy)
|
38 |
+
# Remove \displaystyle
|
39 |
+
sympy = sympy.replace(r'\displaystyle', ' ', -1)
|
40 |
+
# Remove \quad
|
41 |
+
sympy = sympy.replace(r'\quad', ' ', -1).replace(r'\qquad', ' ', -1).replace(r'~', ' ', -1).replace(r'\,', ' ', -1)
|
42 |
+
# Remove $
|
43 |
+
sympy = sympy.replace(r'$', ' ', -1)
|
44 |
+
|
45 |
+
# variable values
|
46 |
+
global VARIABLE_VALUES
|
47 |
+
if len(variable_values) > 0:
|
48 |
+
VARIABLE_VALUES = variable_values
|
49 |
+
else:
|
50 |
+
VARIABLE_VALUES = {}
|
51 |
+
|
52 |
+
# setup listener
|
53 |
+
matherror = MathErrorListener(sympy)
|
54 |
+
|
55 |
+
# stream input
|
56 |
+
stream = InputStream(sympy)
|
57 |
+
lex = PSLexer(stream)
|
58 |
+
lex.removeErrorListeners()
|
59 |
+
lex.addErrorListener(matherror)
|
60 |
+
|
61 |
+
tokens = CommonTokenStream(lex)
|
62 |
+
parser = PSParser(tokens)
|
63 |
+
|
64 |
+
# remove default console error listener
|
65 |
+
parser.removeErrorListeners()
|
66 |
+
parser.addErrorListener(matherror)
|
67 |
+
|
68 |
+
# process the input
|
69 |
+
return_data = None
|
70 |
+
math = parser.math()
|
71 |
+
|
72 |
+
# if a list
|
73 |
+
if math.relation_list():
|
74 |
+
return_data = []
|
75 |
+
|
76 |
+
# go over list items
|
77 |
+
relation_list = math.relation_list().relation_list_content()
|
78 |
+
for list_item in relation_list.relation():
|
79 |
+
expr = convert_relation(list_item)
|
80 |
+
return_data.append(expr)
|
81 |
+
|
82 |
+
# if not, do default
|
83 |
+
else:
|
84 |
+
relation = math.relation()
|
85 |
+
return_data = convert_relation(relation)
|
86 |
+
|
87 |
+
return return_data
|
88 |
+
|
89 |
+
|
90 |
+
def math_answer_cleaning(answer, dataset_name):
|
91 |
+
"""
|
92 |
+
remove irrelevant strings and unify the answer format before checking whether the answers are equal
|
93 |
+
"""
|
94 |
+
def _is_completely_wrapped_by_text(input_string):
|
95 |
+
pattern = r'^\\text{(.*)}$'
|
96 |
+
match = re.match(pattern, input_string)
|
97 |
+
if match:
|
98 |
+
## input_string is completely wrapped by \text{}
|
99 |
+
extracted_content = match.group(1)
|
100 |
+
extracted_content = extracted_content.replace("(", "").replace(")", "").replace(",", "")
|
101 |
+
return extracted_content
|
102 |
+
else:
|
103 |
+
return None
|
104 |
+
|
105 |
+
## remove irrelevant \\text and space
|
106 |
+
extracted_content = _is_completely_wrapped_by_text(answer)
|
107 |
+
answer = extracted_content if extracted_content else answer
|
108 |
+
|
109 |
+
## e.g., convert 5,\!460 into 5460; convert 14{,}916 into 14916 convert \$4 into 4
|
110 |
+
answer = answer.replace(",\!", "").replace("{,}", "").replace("\$", "")
|
111 |
+
## e.g., convert \dfrac{3}{2} into frac{3}{2}
|
112 |
+
answer = answer.replace("dfrac{", "frac{").replace("tfrac{", "frac{")
|
113 |
+
## e.g., convert 121^\circ into 121
|
114 |
+
answer = answer.replace("^\circ", "")
|
115 |
+
answer = answer.replace("^{\circ}", "")
|
116 |
+
## remove \quad
|
117 |
+
answer = answer.replace("\quad", "")
|
118 |
+
## remove space
|
119 |
+
answer = answer.replace(" ", "")
|
120 |
+
## remove \n
|
121 |
+
answer = answer.replace("\n", "").replace("\\n", "")
|
122 |
+
## e.g., convert 3.54\times10^{10} into 3.54e10
|
123 |
+
answer = re.sub(r'([+-]?\d*\.?\d+)[\\]times10\^{([+-]?\d+)}', r'\1e\2', answer)
|
124 |
+
## e.g., convert 3.54\times10^10 into 3.54e10
|
125 |
+
answer = re.sub(r'([+-]?\d*\.?\d+)[\\]times10\^([+-]?\d+)', r'\1e\2', answer)
|
126 |
+
## e.g., convert 558\,\text{nm} into 558
|
127 |
+
answer = re.sub(r'\\,\\text\{.*?\}', '', answer)
|
128 |
+
## e.g., convert 558\text{nm} into 558
|
129 |
+
answer = re.sub(r'\\text\{.*?\}', '', answer)
|
130 |
+
## e.g., convert 2^{10} into 2^10
|
131 |
+
answer = re.sub(r'(\d+)\^{(\d+)}', r'\1^\2', answer)
|
132 |
+
## lowercase
|
133 |
+
answer = answer.lower()
|
134 |
+
|
135 |
+
if dataset_name == "collegemath":
|
136 |
+
## convert 558\mathrm{ft} into 558
|
137 |
+
answer = re.sub(r'\\mathrm\{.*?\}', '', answer)
|
138 |
+
## clean noisy answer
|
139 |
+
answer = re.sub(r'\$\([^)]*\)', '', answer)
|
140 |
+
if answer.endswith("-"):
|
141 |
+
answer = answer[:-1]
|
142 |
+
if answer.endswith("."):
|
143 |
+
answer = answer[:-1]
|
144 |
+
if answer.endswith("hours"):
|
145 |
+
answer = answer[:-len("hours")]
|
146 |
+
## extract final answer after '=' or ':'
|
147 |
+
if "=" in answer:
|
148 |
+
answer = answer.split("=", 1)[1]
|
149 |
+
if ":" in answer:
|
150 |
+
answer = answer.split(":", 1)[1]
|
151 |
+
## \emptyset and \oslash both reprsent empty set in latex
|
152 |
+
answer = answer.replace("\\emptyset", "\\oslash")
|
153 |
+
if dataset_name == "gsm8k":
|
154 |
+
# Example: 5,600 -> 5600
|
155 |
+
answer = answer.replace(',', '')
|
156 |
+
if dataset_name == "gaokao2023en":
|
157 |
+
unit_strings = ['students', 'dollars', 'boxes', 'feet', 'kilometers', 'meters', 'degreesontheBreadusscale', '$', 'a.m.', 'am', 'minutes']
|
158 |
+
for unit in unit_strings:
|
159 |
+
answer = answer.replace(unit, "")
|
160 |
+
|
161 |
+
return answer
|
162 |
+
|
163 |
+
|
164 |
+
def extract_final_answer(output):
|
165 |
+
pattern_re = re.compile(r"\\boxed\{((?:[^{}]|\{(?:[^{}]|\{[^{}]*\})*\})*)\}", re.DOTALL)
|
166 |
+
all_matches = pattern_re.findall(output)
|
167 |
+
|
168 |
+
if len(all_matches) >= 1:
|
169 |
+
extracted_answer = all_matches[-1]
|
170 |
+
else:
|
171 |
+
extracted_answer = None
|
172 |
+
|
173 |
+
return extracted_answer, all_matches
|
174 |
+
|
175 |
+
|
176 |
+
def round_number(answer):
|
177 |
+
def _is_float(string):
|
178 |
+
try:
|
179 |
+
float(string)
|
180 |
+
return True
|
181 |
+
except:
|
182 |
+
return False
|
183 |
+
|
184 |
+
if _is_float(answer) and float(answer) < 1:
|
185 |
+
## to consider the case like 5.56e-10 (convert 5.56e-10 into 5.6e-10)
|
186 |
+
## still return a string type
|
187 |
+
return f"{float(answer):.2g}"
|
188 |
+
|
189 |
+
return answer
|
190 |
+
|
191 |
+
|
192 |
+
def choice_answer_clean(pred: str):
|
193 |
+
pred = pred.strip("\n").rstrip(".").rstrip("/").strip(" ").lstrip(":")
|
194 |
+
# Clean the answer based on the dataset
|
195 |
+
tmp = re.findall(r"\b(A|B|C|D|E)\b", pred.upper())
|
196 |
+
if tmp:
|
197 |
+
pred = tmp
|
198 |
+
else:
|
199 |
+
pred = [pred.strip().strip(".")]
|
200 |
+
pred = pred[-1]
|
201 |
+
# Remove the period at the end, again!
|
202 |
+
pred = pred.rstrip(".").rstrip("/")
|
203 |
+
return pred
|
204 |
+
|
205 |
+
|
206 |
+
def parse_digits(num):
|
207 |
+
num = regex.sub(",", "", str(num))
|
208 |
+
try:
|
209 |
+
return float(num)
|
210 |
+
except:
|
211 |
+
if num.endswith("%"):
|
212 |
+
num = num[:-1]
|
213 |
+
if num.endswith("\\"):
|
214 |
+
num = num[:-1]
|
215 |
+
try:
|
216 |
+
return float(num) / 100
|
217 |
+
except:
|
218 |
+
pass
|
219 |
+
return None
|
220 |
+
|
221 |
+
|
222 |
+
def is_digit(num):
|
223 |
+
# paired with parse_digits
|
224 |
+
return parse_digits(num) is not None
|
225 |
+
|
226 |
+
|
227 |
+
def str_to_pmatrix(input_str):
|
228 |
+
input_str = input_str.strip()
|
229 |
+
matrix_str = re.findall(r"\{.*,.*\}", input_str)
|
230 |
+
pmatrix_list = []
|
231 |
+
|
232 |
+
for m in matrix_str:
|
233 |
+
m = m.strip("{}")
|
234 |
+
pmatrix = r"\begin{pmatrix}" + m.replace(",", "\\") + r"\end{pmatrix}"
|
235 |
+
pmatrix_list.append(pmatrix)
|
236 |
+
|
237 |
+
return ", ".join(pmatrix_list)
|
238 |
+
|
239 |
+
|
240 |
+
def math_equal(
|
241 |
+
prediction: Union[bool, float, str],
|
242 |
+
reference: Union[float, str],
|
243 |
+
include_percentage: bool = True,
|
244 |
+
is_close: bool = True,
|
245 |
+
timeout: bool = False,
|
246 |
+
) -> bool:
|
247 |
+
"""
|
248 |
+
Exact match of math if and only if:
|
249 |
+
1. numerical equal: both can convert to float and are equal
|
250 |
+
2. symbolic equal: both can convert to sympy expression and are equal
|
251 |
+
"""
|
252 |
+
if prediction is None or reference is None:
|
253 |
+
return False
|
254 |
+
if str(prediction.strip().lower()) == str(reference.strip().lower()):
|
255 |
+
return True
|
256 |
+
if (
|
257 |
+
reference in ["A", "B", "C", "D", "E"]
|
258 |
+
and choice_answer_clean(prediction) == reference
|
259 |
+
):
|
260 |
+
return True
|
261 |
+
|
262 |
+
# fraction equal
|
263 |
+
if fraction_equal(prediction, reference):
|
264 |
+
return True
|
265 |
+
|
266 |
+
try: # numerical equal
|
267 |
+
if round_number(prediction) == round_number(reference):
|
268 |
+
return True
|
269 |
+
if is_digit(prediction) and is_digit(reference):
|
270 |
+
prediction = parse_digits(prediction)
|
271 |
+
reference = parse_digits(reference)
|
272 |
+
# number questions
|
273 |
+
if include_percentage:
|
274 |
+
gt_result = [reference / 100, reference, reference * 100]
|
275 |
+
else:
|
276 |
+
gt_result = [reference]
|
277 |
+
for item in gt_result:
|
278 |
+
try:
|
279 |
+
if is_close:
|
280 |
+
if numeric_equal(prediction, item):
|
281 |
+
return True
|
282 |
+
else:
|
283 |
+
if item == prediction:
|
284 |
+
return True
|
285 |
+
except Exception:
|
286 |
+
continue
|
287 |
+
return False
|
288 |
+
except:
|
289 |
+
pass
|
290 |
+
|
291 |
+
if not prediction and prediction not in [0, False]:
|
292 |
+
return False
|
293 |
+
|
294 |
+
# symbolic equal
|
295 |
+
reference = str(reference).strip()
|
296 |
+
prediction = str(prediction).strip()
|
297 |
+
|
298 |
+
## pmatrix (amps)
|
299 |
+
if "pmatrix" in prediction and not "pmatrix" in reference:
|
300 |
+
reference = str_to_pmatrix(reference)
|
301 |
+
|
302 |
+
## deal with [], (), {}
|
303 |
+
pred_str, ref_str = prediction, reference
|
304 |
+
if (
|
305 |
+
prediction.startswith("[")
|
306 |
+
and prediction.endswith("]")
|
307 |
+
and not reference.startswith("(")
|
308 |
+
) or (
|
309 |
+
prediction.startswith("(")
|
310 |
+
and prediction.endswith(")")
|
311 |
+
and not reference.startswith("[")
|
312 |
+
):
|
313 |
+
pred_str = pred_str.strip("[]()")
|
314 |
+
ref_str = ref_str.strip("[]()")
|
315 |
+
for s in ["{", "}", "(", ")"]:
|
316 |
+
ref_str = ref_str.replace(s, "")
|
317 |
+
pred_str = pred_str.replace(s, "")
|
318 |
+
if pred_str.lower() == ref_str.lower():
|
319 |
+
return True
|
320 |
+
|
321 |
+
## [a, b] vs. [c, d], return a==c and b==d
|
322 |
+
if (
|
323 |
+
regex.match(r"(\(|\[).+(\)|\])", prediction) is not None
|
324 |
+
and regex.match(r"(\(|\[).+(\)|\])", reference) is not None
|
325 |
+
):
|
326 |
+
pred_parts = prediction[1:-1].split(",")
|
327 |
+
ref_parts = reference[1:-1].split(",")
|
328 |
+
if len(pred_parts) == len(ref_parts):
|
329 |
+
if all(
|
330 |
+
[
|
331 |
+
math_equal(
|
332 |
+
pred_parts[i], ref_parts[i], include_percentage, is_close
|
333 |
+
)
|
334 |
+
for i in range(len(pred_parts))
|
335 |
+
]
|
336 |
+
):
|
337 |
+
return True
|
338 |
+
if (
|
339 |
+
(
|
340 |
+
prediction.startswith("\\begin{pmatrix}")
|
341 |
+
or prediction.startswith("\\begin{bmatrix}")
|
342 |
+
)
|
343 |
+
and (
|
344 |
+
prediction.endswith("\\end{pmatrix}")
|
345 |
+
or prediction.endswith("\\end{bmatrix}")
|
346 |
+
)
|
347 |
+
and (
|
348 |
+
reference.startswith("\\begin{pmatrix}")
|
349 |
+
or reference.startswith("\\begin{bmatrix}")
|
350 |
+
)
|
351 |
+
and (
|
352 |
+
reference.endswith("\\end{pmatrix}") or reference.endswith("\\end{bmatrix}")
|
353 |
+
)
|
354 |
+
):
|
355 |
+
pred_lines = [
|
356 |
+
line.strip()
|
357 |
+
for line in prediction[
|
358 |
+
len("\\begin{pmatrix}") : -len("\\end{pmatrix}")
|
359 |
+
].split("\\\\")
|
360 |
+
if line.strip()
|
361 |
+
]
|
362 |
+
ref_lines = [
|
363 |
+
line.strip()
|
364 |
+
for line in reference[
|
365 |
+
len("\\begin{pmatrix}") : -len("\\end{pmatrix}")
|
366 |
+
].split("\\\\")
|
367 |
+
if line.strip()
|
368 |
+
]
|
369 |
+
matched = True
|
370 |
+
if len(pred_lines) == len(ref_lines):
|
371 |
+
for pred_line, ref_line in zip(pred_lines, ref_lines):
|
372 |
+
pred_parts = pred_line.split("&")
|
373 |
+
ref_parts = ref_line.split("&")
|
374 |
+
if len(pred_parts) == len(ref_parts):
|
375 |
+
if not all(
|
376 |
+
[
|
377 |
+
math_equal(
|
378 |
+
pred_parts[i],
|
379 |
+
ref_parts[i],
|
380 |
+
include_percentage,
|
381 |
+
is_close,
|
382 |
+
)
|
383 |
+
for i in range(len(pred_parts))
|
384 |
+
]
|
385 |
+
):
|
386 |
+
matched = False
|
387 |
+
break
|
388 |
+
else:
|
389 |
+
matched = False
|
390 |
+
if not matched:
|
391 |
+
break
|
392 |
+
else:
|
393 |
+
matched = False
|
394 |
+
if matched:
|
395 |
+
return True
|
396 |
+
|
397 |
+
if prediction.count("=") == 1 and reference.count("=") == 1:
|
398 |
+
pred = prediction.split("=")
|
399 |
+
pred = f"{pred[0].strip()} - ({pred[1].strip()})"
|
400 |
+
ref = reference.split("=")
|
401 |
+
ref = f"{ref[0].strip()} - ({ref[1].strip()})"
|
402 |
+
if symbolic_equal(pred, ref) or symbolic_equal(f"-({pred})", ref):
|
403 |
+
return True
|
404 |
+
elif (
|
405 |
+
prediction.count("=") == 1
|
406 |
+
and len(prediction.split("=")[0].strip()) <= 2
|
407 |
+
and "=" not in reference
|
408 |
+
):
|
409 |
+
if math_equal(
|
410 |
+
prediction.split("=")[1], reference, include_percentage, is_close
|
411 |
+
):
|
412 |
+
return True
|
413 |
+
elif (
|
414 |
+
reference.count("=") == 1
|
415 |
+
and len(reference.split("=")[0].strip()) <= 2
|
416 |
+
and "=" not in prediction
|
417 |
+
):
|
418 |
+
if math_equal(
|
419 |
+
prediction, reference.split("=")[1], include_percentage, is_close
|
420 |
+
):
|
421 |
+
return True
|
422 |
+
|
423 |
+
# symbolic equal with sympy
|
424 |
+
if timeout:
|
425 |
+
if call_with_timeout(symbolic_equal_process, prediction, reference):
|
426 |
+
return True
|
427 |
+
else:
|
428 |
+
if symbolic_equal(prediction, reference):
|
429 |
+
return True
|
430 |
+
|
431 |
+
return False
|
432 |
+
|
433 |
+
|
434 |
+
def numeric_equal(prediction: float, reference: float):
|
435 |
+
# Note that relative tolerance has significant impact
|
436 |
+
# on the result of the synthesized GSM-Hard dataset
|
437 |
+
# if reference.is_integer():
|
438 |
+
# return isclose(reference, round(prediction), abs_tol=1e-4)
|
439 |
+
# else:
|
440 |
+
# prediction = round(prediction, len(str(reference).split(".")[-1]))
|
441 |
+
return isclose(reference, prediction, rel_tol=1e-4)
|
442 |
+
|
443 |
+
|
444 |
+
def fraction_equal(prediction, reference):
|
445 |
+
def _calculate_numbers(input_string):
|
446 |
+
try:
|
447 |
+
result = eval(input_string)
|
448 |
+
return result
|
449 |
+
except:
|
450 |
+
return None
|
451 |
+
|
452 |
+
reference = re.sub(r'\\frac{(.*?)}{(.*?)}', r'(\1/\2)', reference)
|
453 |
+
prediction = re.sub(r'\\frac{(.*?)}{(.*?)}', r'(\1/\2)', prediction)
|
454 |
+
|
455 |
+
if reference == prediction:
|
456 |
+
return True
|
457 |
+
|
458 |
+
reference = _calculate_numbers(reference)
|
459 |
+
prediction = _calculate_numbers(prediction)
|
460 |
+
|
461 |
+
if reference and reference == prediction:
|
462 |
+
return True
|
463 |
+
|
464 |
+
return False
|
465 |
+
|
466 |
+
def symbolic_equal(a, b):
|
467 |
+
def _parse(s):
|
468 |
+
for f in [parse_latex, parse_expr, latex2sympy]:
|
469 |
+
try:
|
470 |
+
return f(s.replace("\\\\", "\\"))
|
471 |
+
except:
|
472 |
+
try:
|
473 |
+
return f(s)
|
474 |
+
except:
|
475 |
+
pass
|
476 |
+
return s
|
477 |
+
|
478 |
+
a = _parse(a)
|
479 |
+
b = _parse(b)
|
480 |
+
|
481 |
+
# direct equal
|
482 |
+
try:
|
483 |
+
if str(a) == str(b) or a == b:
|
484 |
+
return True
|
485 |
+
except:
|
486 |
+
pass
|
487 |
+
|
488 |
+
# simplify equal
|
489 |
+
try:
|
490 |
+
if a.equals(b) or simplify(a - b) == 0:
|
491 |
+
return True
|
492 |
+
except:
|
493 |
+
pass
|
494 |
+
|
495 |
+
# equation equal
|
496 |
+
try:
|
497 |
+
if (abs(a.lhs - a.rhs)).equals(abs(b.lhs - b.rhs)):
|
498 |
+
return True
|
499 |
+
except:
|
500 |
+
pass
|
501 |
+
|
502 |
+
try:
|
503 |
+
if numeric_equal(float(N(a)), float(N(b))):
|
504 |
+
return True
|
505 |
+
except:
|
506 |
+
pass
|
507 |
+
|
508 |
+
# matrix
|
509 |
+
try:
|
510 |
+
# if a and b are matrix
|
511 |
+
if a.shape == b.shape:
|
512 |
+
_a = a.applyfunc(lambda x: round(x, 3))
|
513 |
+
_b = b.applyfunc(lambda x: round(x, 3))
|
514 |
+
if _a.equals(_b):
|
515 |
+
return True
|
516 |
+
except:
|
517 |
+
pass
|
518 |
+
|
519 |
+
return False
|
520 |
+
|
521 |
+
|
522 |
+
def symbolic_equal_process(a, b, output_queue):
|
523 |
+
result = symbolic_equal(a, b)
|
524 |
+
output_queue.put(result)
|
525 |
+
|
526 |
+
|
527 |
+
def math_equal_process(prediction, reference, output_queue):
|
528 |
+
result = math_equal(prediction, reference, timeout=True)
|
529 |
+
output_queue.put(result)
|
530 |
+
|
531 |
+
|
532 |
+
def call_with_timeout(func, *args, timeout=1, **kwargs):
|
533 |
+
output_queue = multiprocessing.Queue()
|
534 |
+
process_args = args + (output_queue,)
|
535 |
+
process = multiprocessing.Process(target=func, args=process_args, kwargs=kwargs)
|
536 |
+
process.start()
|
537 |
+
process.join(timeout)
|
538 |
+
|
539 |
+
if process.is_alive():
|
540 |
+
process.terminate()
|
541 |
+
process.join()
|
542 |
+
return False
|
543 |
+
|
544 |
+
return output_queue.get()
|
545 |
+
|
546 |
+
|
547 |
+
def check_correctness_of_multiple_answer_cases(prediction, reference, all_matches):
|
548 |
+
|
549 |
+
if prediction.replace(",", "").replace("$", "") == reference.replace(",", "").replace("$", ""):
|
550 |
+
return True
|
551 |
+
|
552 |
+
if not prediction.split("=")[-1] == reference.split("=")[-1].replace("$", ""):
|
553 |
+
return False
|
554 |
+
|
555 |
+
if "," in reference or "or" in reference or "and" in reference:
|
556 |
+
## there are multiple answers
|
557 |
+
if len(all_matches) <= 1:
|
558 |
+
return False
|
559 |
+
|
560 |
+
prediction1 = prediction.split("=")[-1]
|
561 |
+
prediction2 = all_matches[-2].split("=")[-1]
|
562 |
+
reference = reference.replace("$", "")
|
563 |
+
if "or" in reference:
|
564 |
+
gold_list = reference.split("or", 1)
|
565 |
+
elif "and" in reference:
|
566 |
+
gold_list = reference.split("and", 1)
|
567 |
+
else:
|
568 |
+
gold_list = reference.split(",", 1)
|
569 |
+
|
570 |
+
reference1 = gold_list[-1].split("=")[-1]
|
571 |
+
reference2 = gold_list[-2].split("=")[-1]
|
572 |
+
|
573 |
+
if math_equal(prediction1, reference1) and math_equal(prediction2, reference2):
|
574 |
+
return True
|
575 |
+
elif math_equal(prediction2, reference1) and math_equal(prediction1, reference2):
|
576 |
+
return True
|
577 |
+
|
578 |
+
return False
|
579 |
+
|
580 |
+
else:
|
581 |
+
return True
|
582 |
+
|
583 |
+
|
584 |
+
def is_equal(model_output, reference, dataset_name):
|
585 |
+
|
586 |
+
extracted_model_answer, all_matches = extract_final_answer(model_output)
|
587 |
+
if extracted_model_answer is None or reference is None:
|
588 |
+
return False
|
589 |
+
|
590 |
+
extracted_model_answer = math_answer_cleaning(extracted_model_answer, dataset_name)
|
591 |
+
reference = math_answer_cleaning(reference, dataset_name)
|
592 |
+
|
593 |
+
# if math_equal(prediction, reference, timeout=True):
|
594 |
+
if call_with_timeout(math_equal_process, extracted_model_answer, reference):
|
595 |
+
return True
|
596 |
+
|
597 |
+
if dataset_name == "collegemath":
|
598 |
+
return check_correctness_of_multiple_answer_cases(extracted_model_answer, reference, all_matches)
|
599 |
+
|
600 |
+
return False
|