Update mathrandom-progressive.py
Browse files- mathrandom-progressive.py +134 -23
mathrandom-progressive.py
CHANGED
@@ -1,46 +1,157 @@
|
|
1 |
import random
|
2 |
import json
|
3 |
import operator
|
|
|
4 |
|
5 |
num_samples = [20000, 50000, 230000, 500000]
|
6 |
min_values = [-9.99, -19.99, -99.99, -99.99]
|
7 |
max_values = [9.99, 19.99, 99.99, 99.99]
|
8 |
vals_floats = ["%.2f", "%.3f", "%.3f", "%.4f"]
|
9 |
result_floats = ["%.2f", "%.3f", "%.4f", "%.4f"]
|
10 |
-
|
|
|
|
|
|
|
|
|
|
|
11 |
'+': operator.add,
|
12 |
'-': operator.sub,
|
13 |
'*': operator.mul,
|
14 |
-
'/': operator.truediv
|
|
|
|
|
15 |
}
|
16 |
|
|
|
17 |
def safe_division(numerator, denominator):
|
18 |
return numerator if denominator == 0 else numerator / denominator
|
19 |
|
20 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
data = []
|
22 |
-
for _ in range(num_samples
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
if
|
33 |
-
|
34 |
-
|
35 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
36 |
else:
|
37 |
-
|
38 |
-
|
39 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
40 |
|
41 |
-
|
42 |
-
data.append({'instruction': expression, 'output': output})
|
43 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
44 |
out_file = f'arithmetic-float-complex_{idx}.json'
|
45 |
with open(out_file, 'w') as f:
|
46 |
-
json.dump(data, f)
|
|
|
|
|
|
|
|
1 |
import random
|
2 |
import json
|
3 |
import operator
|
4 |
+
import math
|
5 |
|
6 |
num_samples = [20000, 50000, 230000, 500000]
|
7 |
min_values = [-9.99, -19.99, -99.99, -99.99]
|
8 |
max_values = [9.99, 19.99, 99.99, 99.99]
|
9 |
vals_floats = ["%.2f", "%.3f", "%.3f", "%.4f"]
|
10 |
result_floats = ["%.2f", "%.3f", "%.4f", "%.4f"]
|
11 |
+
seed = 42
|
12 |
+
|
13 |
+
random.seed(seed)
|
14 |
+
|
15 |
+
# Binary operations
|
16 |
+
binary_operations = {
|
17 |
'+': operator.add,
|
18 |
'-': operator.sub,
|
19 |
'*': operator.mul,
|
20 |
+
'/': operator.truediv,
|
21 |
+
#'**': operator.pow, # it makes tons of troubles, but if u wanna contribute with a functional fix (L)
|
22 |
+
'%': operator.mod
|
23 |
}
|
24 |
|
25 |
+
# Safe division to avoid division by zero
|
26 |
def safe_division(numerator, denominator):
|
27 |
return numerator if denominator == 0 else numerator / denominator
|
28 |
|
29 |
+
# Unary operations
|
30 |
+
def safe_sqrt(x):
|
31 |
+
return math.sqrt(x) if x >= 0 else 0
|
32 |
+
|
33 |
+
def safe_log(x):
|
34 |
+
return math.log(x) if x > 0 else 0
|
35 |
+
|
36 |
+
unary_operations = {
|
37 |
+
'sqrt': safe_sqrt,
|
38 |
+
'sin': math.sin,
|
39 |
+
'cos': math.cos,
|
40 |
+
'log': safe_log
|
41 |
+
}
|
42 |
+
|
43 |
+
def format_result(result, format_spec):
|
44 |
+
if isinstance(result, complex):
|
45 |
+
real_part = format_spec % result.real
|
46 |
+
imag_part = format_spec % result.imag
|
47 |
+
return f"{real_part} + {imag_part}j"
|
48 |
+
else:
|
49 |
+
return format_spec % result
|
50 |
+
|
51 |
+
def safe_modulo(numerator, denominator):
|
52 |
+
return 0 if denominator == 0 else numerator % denominator
|
53 |
+
|
54 |
+
def safe_power(base, exponent):
|
55 |
+
if base == 0 and exponent < 0:
|
56 |
+
return 0 # or some other value that signifies an undefined operation
|
57 |
+
else:
|
58 |
+
return base ** exponent
|
59 |
+
|
60 |
+
def generate_sub_expression(min_value, max_value, val_float):
|
61 |
+
num1 = float(val_float % random.uniform(min_value, max_value))
|
62 |
+
num2 = float(val_float % random.uniform(min_value, max_value))
|
63 |
+
operation = random.choice(list(binary_operations.keys()))
|
64 |
+
|
65 |
+
if operation == '/':
|
66 |
+
result = safe_division(num1, num2)
|
67 |
+
elif operation == '%':
|
68 |
+
result = safe_modulo(num1, num2)
|
69 |
+
else:
|
70 |
+
result = binary_operations[operation](num1, num2)
|
71 |
+
|
72 |
+
expression = f"({num1} {operation} {num2})"
|
73 |
+
return expression, result
|
74 |
+
|
75 |
+
def generate_expression(num_samples, min_value, max_value, val_float, result_float):
|
76 |
data = []
|
77 |
+
for _ in range(num_samples):
|
78 |
+
sub_ops_count = random.randint(0, 2) # Number of sub-operations (0, 1, or 2), the code is not tailored for more cases, but feel free to improve it.
|
79 |
+
sub_expressions = [generate_sub_expression(min_value, max_value, val_float) for _ in range(sub_ops_count)]
|
80 |
+
|
81 |
+
# Main operation
|
82 |
+
main_num1 = float(val_float % random.uniform(min_value, max_value))
|
83 |
+
main_num2 = float(val_float % random.uniform(min_value, max_value))
|
84 |
+
main_operation = random.choice(list(binary_operations.keys()))
|
85 |
+
|
86 |
+
# Embed sub-expressions into main expression
|
87 |
+
if sub_expressions:
|
88 |
+
if len(sub_expressions) == 1:
|
89 |
+
sub_expr, sub_result = sub_expressions[0]
|
90 |
+
if random.choice([True, False]):
|
91 |
+
main_expression = f"{sub_expr} {main_operation} {main_num2}"
|
92 |
+
if main_operation == '/':
|
93 |
+
main_result = safe_division(sub_result, main_num2)
|
94 |
+
elif main_operation == '%':
|
95 |
+
main_result = safe_modulo(sub_result, main_num2)
|
96 |
+
else:
|
97 |
+
main_result = binary_operations[main_operation](sub_result, main_num2)
|
98 |
+
else:
|
99 |
+
main_expression = f"{main_num1} {main_operation} {sub_expr}"
|
100 |
+
if main_operation == '/':
|
101 |
+
main_result = safe_division(main_num1, sub_result)
|
102 |
+
elif main_operation == '%':
|
103 |
+
main_result = safe_modulo(main_num1, sub_result)
|
104 |
+
else:
|
105 |
+
main_result = binary_operations[main_operation](main_num1, sub_result)
|
106 |
+
else:
|
107 |
+
sub_expr1, sub_result1 = sub_expressions[0]
|
108 |
+
sub_expr2, sub_result2 = sub_expressions[1]
|
109 |
+
main_expression = f"{sub_expr1} {main_operation} {sub_expr2}"
|
110 |
+
if main_operation == '/':
|
111 |
+
main_result = safe_division(sub_result1, sub_result2)
|
112 |
+
elif main_operation == '%':
|
113 |
+
main_result = safe_modulo(sub_result1, sub_result2)
|
114 |
+
else:
|
115 |
+
main_result = binary_operations[main_operation](sub_result1, sub_result2)
|
116 |
else:
|
117 |
+
main_expression = f"{main_num1} {main_operation} {main_num2}"
|
118 |
+
if main_operation == '/':
|
119 |
+
main_result = safe_division(main_num1, main_num2)
|
120 |
+
elif main_operation == '%':
|
121 |
+
main_result = safe_modulo(main_num1, main_num2)
|
122 |
+
else:
|
123 |
+
main_result = binary_operations[main_operation](main_num1, main_num2)
|
124 |
+
|
125 |
+
output = format_result(main_result, result_float)
|
126 |
+
data.append({'instruction': main_expression, 'output': output})
|
127 |
|
128 |
+
return data
|
|
|
129 |
|
130 |
+
|
131 |
+
# DPO (Follows Ultrabinarized Format)
|
132 |
+
def make_prompt(expression, output):
|
133 |
+
prompt = {
|
134 |
+
'prompt': 'What is the result of the following expression?',
|
135 |
+
'chosen': [],
|
136 |
+
'rejected': [],
|
137 |
+
'messages': []
|
138 |
+
}
|
139 |
+
prompt['chosen'].append({'content': prompt['prompt'] + '\n' + expression, 'role': 'user'})
|
140 |
+
prompt['chosen'].append({'content': output, 'role': 'assistant'})
|
141 |
+
prompt['rejected'].append({'content': prompt['prompt'] + '\n' + expression, 'role': 'user'})
|
142 |
+
prompt['rejected'].append({'content': output[0:2], 'role': 'assistant'})
|
143 |
+
prompt['messages'].append({'content': prompt['prompt'] + '\n' + expression, 'role': 'user'})
|
144 |
+
prompt['messages'].append({'content': output, 'role': 'assistant'})
|
145 |
+
return prompt
|
146 |
+
|
147 |
+
|
148 |
+
# Generate and write data for each set of parameters
|
149 |
+
for idx, sample_count in enumerate(num_samples):
|
150 |
+
data = generate_expression(sample_count, min_values[idx], max_values[idx], vals_floats[idx], result_floats[idx])
|
151 |
+
dpo_data = [make_prompt(elem['instruction'], elem['output']) for elem in data]
|
152 |
out_file = f'arithmetic-float-complex_{idx}.json'
|
153 |
with open(out_file, 'w') as f:
|
154 |
+
json.dump(data, f)
|
155 |
+
out_file = f'arithmetic-float-complex_DPO_{idx}.json'
|
156 |
+
with open(out_file, 'w') as f:
|
157 |
+
json.dump(dpo_data, f)
|