File size: 17,720 Bytes
a8f16a9 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 |
import random
import json
import math
import time
import re
import sys
import multiprocessing
import os
from tqdm import tqdm
NUM_LINES = 2000000
OUTPUT_FILE = "correct_math_data.jsonl"
MIN_LENGTH = 2
MAX_LENGTH = 8
MIN_NUMBER = 1
MAX_NUMBER = 999
MAX_EXPONENT_BASE = 9
MAX_EXPONENT_POWER = 5
REASONING_CHANCE = 0.8
WORD_FORM_CHANCE = 0.25
BRACKET_CHANCE = 0.5
SENTENCE_FORM_CHANCE = 0.6
MAX_SOLVER_ITERATIONS = 30 # Reduced from 50 for faster timeout
NUM_WORKERS = os.cpu_count() or 1
PROMPT_TEMPLATES = [
"What is {expression}?", "Calculate the value of {expression}.", "Find the result of {expression}.",
"Can you solve {expression}?", "Solve for {expression}.", "What does {expression} equal?", "Compute {expression}.",
"What is the solution to {expression}?", "Give me the answer for {expression}.", "Determine the value of {expression}.",
"Evaluate the expression: {expression}.", "I need the result of {expression}, please."
]
COT_INTRO_TEMPLATES = [
"<think> Let's break down the equation {expression} step by step, following the order of operations (BEDMAS).",
"<think> Okay, to solve {expression}, I'll follow BEDMAS (Brackets, Exponents, Division/Multiplication, Addition/Subtraction).",
"<think> Analyzing {expression}. I need to solve this by applying the correct order of operations.",
"<think> Here's my step-by-step evaluation for {expression}:",
"<think> To get the answer for {expression}, I will use the order of operations.",
"<think> Processing {expression} requires following BEDMAS, let's begin.",
"<think> I will solve {expression} by carefully following the rules of BEDMAS.",
"<think> The expression is {expression}. My plan is to solve it using the order of operations.",
"<think> To solve this, I'll go through Brackets, then Exponents, then Multiplication/Division, and finally Addition/Subtraction for {expression}.",
"<think> Let's start solving {expression}. I'll tackle it one operation at a time based on BEDMAS.",
"<think> Thinking step-by-step for {expression}..."
]
COT_STEP_TEMPLATES = {
"brackets": [
"First, I'll solve the expression inside the brackets: {part}. That equals {result}.",
"Starting with the parentheses, {part} evaluates to {result}.",
"The brackets are the priority. Calculating {part} gives me {result}.",
"The calculation inside the parentheses comes first: {part} becomes {result}.",
"Looking inside the brackets, I see {part}. The result of that is {result}.",
"I'll begin by simplifying the part in the parentheses: {part} is {result}.",
"The first step according to BEDMAS is brackets. So, {part} is solved to {result}.",
"Tackling the parentheses first: {part} simplifies to {result}.",
"Evaluating the bracketed expression {part} yields {result}.",
"My focus is on the brackets first. {part} equals {result}."
],
"exponents": [
"Next, I'll handle the exponents. {part} is {result}.",
"Exponents are next in order. {part} calculates to {result}.",
"Now for the powers: {part} equals {result}.",
"Moving on to exponents, {part} results in {result}.",
"The next priority is exponents. The term {part} becomes {result}.",
"After brackets, I solve for exponents. {part} gives {result}.",
"Now, calculating the power: {part} is equal to {result}.",
"I see an exponent at {part}. This evaluates to {result}.",
"The 'E' in BEDMAS is for exponents, so I'll solve {part} to get {result}.",
"Time to resolve the exponents. {part} is {result}."
],
"multi_div_mod": [
"Now, I'll perform multiplication, division, and modulo from left to right. The first is {part}, which is {result}.",
"Next up is multiplication and division. I see {part}, which gives {result}.",
"Working through multiplication/division from left to right, {part} results in {result}.",
"The next step is to resolve multiplication and division. {part} is {result}.",
"Scanning from left to right for M/D/M, I find {part}. This calculates to {result}.",
"Now for multiplication and division. The operation {part} equals {result}.",
"Moving on, I'll handle the multiplication/division. {part} becomes {result}.",
"The next operations are multiply and divide. I'll solve {part} to get {result}.",
"I will now compute {part}, which results in {result}.",
"Left-to-right, the next multiplication or division is {part}, giving {result}."
],
"add_sub": [
"Finally, I'll do the addition and subtraction from left to right. I have {part}, which equals {result}.",
"Last step is addition and subtraction. {part} becomes {result}.",
"Finishing up with addition/subtraction, {part} evaluates to {result}.",
"The final operations are addition and subtraction. {part} results in {result}.",
"Now for the final calculations, addition and subtraction. {part} is {result}.",
"Working from left to right, the final step is {part}, which is {result}.",
"The last part of BEDMAS is addition and subtraction. {part} gives {result}.",
"To finish, I'll solve {part}, resulting in {result}.",
"Finally, the addition/subtraction part: {part} equals {result}.",
"The last calculation is {part}, and the answer is {result}."
]
}
COT_FINALIZER_TEMPLATES = [
"After all steps, the final answer is {result}.",
"So, the complete result for the expression is {result}.",
"Therefore, the final value is {result}.",
"Bringing it all together, the answer is {result}.",
"The final computation yields {result}.",
"Thus, the expression evaluates to {result}.",
"So the final answer is {result}.",
"After all those steps, we arrive at the answer: {result}.",
"The result of the entire calculation is {result}.",
"In conclusion, the answer is {result}."
]
SIMPLE_COMPLETION_TEMPLATES = [
"The equation {expression} equals {result}.", "The answer is {result}.",
"The result is {result}.", "It equals {result}.", "The final value is {result}.",
"{expression} results in {result}.", "The solution is {result}.",
"The value is {result}.", "After calculation, the answer is {result}.",
"The final result is {result}."
]
ONES = ['', 'one', 'two', 'three', 'four', 'five', 'six', 'seven', 'eight', 'nine']
TENS = ['', '', 'twenty', 'thirty', 'forty', 'fifty', 'sixty', 'seventy', 'eighty', 'ninety']
TEENS = ['ten', 'eleven', 'twelve', 'thirteen', 'fourteen', 'fifteen', 'sixteen', 'seventeen', 'eighteen', 'nineteen']
def number_to_words(n):
if not isinstance(n, int): return str(n)
if n == 0: return 'zero'
if n < 0: return f"negative {number_to_words(abs(n))}"
if n < 10: return ONES[n]
if n < 20: return TEENS[n-10]
if n < 100: return TENS[n//10] + (f"-{ONES[n%10]}" if n%10 else "")
if n < 1000: return f"{ONES[n//100]} hundred" + (f" and {number_to_words(n%100)}" if n%100 else "")
if n < 1000000: return f"{number_to_words(n//1000)} thousand" + (f", {number_to_words(n%1000)}" if n%1000 else "")
return str(n)
def operator_to_word(op):
return {'+': 'plus', '-': 'minus', '*': 'times', '/': 'divided by', '^': 'to the power of', '%': 'modulo'}.get(op, op)
def format_number(n):
if isinstance(n, float) and not n.is_integer():
return f"{n:.4f}".rstrip('0').rstrip('.')
return str(int(round(n)))
def generate_expression_parts():
length = random.randint(MIN_LENGTH, MAX_LENGTH)
parts = []
for i in range(length):
if parts and parts[-1] == '^':
parts.append(random.randint(2, MAX_EXPONENT_POWER))
else:
parts.append(random.randint(MIN_NUMBER, MAX_NUMBER))
if i < length - 1:
if parts and parts[-1] != '^':
op = random.choice(['+', '-', '*', '/', '%', '^'])
else:
op = random.choice(['+', '-', '*', '/', '%'])
if op == '^':
parts[-1] = random.randint(MIN_NUMBER, MAX_EXPONENT_BASE)
parts.append(op)
if random.random() < BRACKET_CHANCE and len(parts) >= 5:
start = random.randrange(0, len(parts) - 2, 2)
end = random.randrange(start + 2, len(parts), 2)
parts.insert(end + 1, ')')
parts.insert(start, '(')
return parts
def solve_with_cot(expression_str):
"""Optimized solver with better pattern matching and guaranteed termination."""
steps = []
current_expr = expression_str.strip()
for iteration in range(MAX_SOLVER_ITERATIONS):
# Remove extra spaces
current_expr = re.sub(r'\s+', ' ', current_expr).strip()
# Check if we're done (single number)
try:
final_result = float(current_expr)
return {'steps': steps, 'result': final_result}
except ValueError:
pass
reduction_made = False
# 1. Handle brackets first
bracket_match = re.search(r'\(([^()]+)\)', current_expr)
if bracket_match:
bracket_content = bracket_match.group(1).strip()
sub_solver_result = solve_with_cot(bracket_content)
if not sub_solver_result:
return None
result = sub_solver_result['result']
try:
formatted_result = format_number(result)
except (ValueError, OverflowError):
return None
steps.append(random.choice(COT_STEP_TEMPLATES["brackets"]).format(part=bracket_content, result=formatted_result))
current_expr = current_expr[:bracket_match.start()] + ' ' + formatted_result + ' ' + current_expr[bracket_match.end():]
reduction_made = True
continue
# 2. Handle exponents
exp_match = re.search(r'(-?\d+(?:\.\d+)?)\s*\^\s*(-?\d+(?:\.\d+)?)', current_expr)
if exp_match:
base_str, exp_str = exp_match.groups()
try:
base = float(base_str)
exponent = float(exp_str)
result = base ** exponent
if abs(result) > 1e12 or math.isnan(result) or math.isinf(result):
return None
formatted_result = format_number(result)
except (OverflowError, ValueError, ZeroDivisionError):
return None
part = f"{base_str} ^ {exp_str}"
steps.append(random.choice(COT_STEP_TEMPLATES["exponents"]).format(part=part, result=formatted_result))
current_expr = current_expr[:exp_match.start()] + ' ' + formatted_result + ' ' + current_expr[exp_match.end():]
reduction_made = True
continue
# 3. Handle multiplication, division, modulo (left to right)
mdm_match = re.search(r'(-?\d+(?:\.\d+)?)\s*([*/%])\s*(-?\d+(?:\.\d+)?)', current_expr)
if mdm_match:
left_str, op, right_str = mdm_match.groups()
try:
left = float(left_str)
right = float(right_str)
if op == '*':
result = left * right
elif op == '/':
if right == 0:
return None
result = left / right
elif op == '%':
if right == 0:
return None
result = left % right
if abs(result) > 1e12 or math.isnan(result) or math.isinf(result):
return None
formatted_result = format_number(result)
except (OverflowError, ValueError, ZeroDivisionError):
return None
part = f"{left_str} {op} {right_str}"
steps.append(random.choice(COT_STEP_TEMPLATES["multi_div_mod"]).format(part=part, result=formatted_result))
current_expr = current_expr[:mdm_match.start()] + ' ' + formatted_result + ' ' + current_expr[mdm_match.end():]
reduction_made = True
continue
# 4. Handle addition and subtraction (left to right)
# Match pattern where we have number [+|-] number but not at start of negative number
as_match = re.search(r'(-?\d+(?:\.\d+)?)\s*([+\-])\s*(-?\d+(?:\.\d+)?)', current_expr)
if as_match:
left_str, op, right_str = as_match.groups()
try:
left = float(left_str)
right = float(right_str)
if op == '+':
result = left + right
elif op == '-':
result = left - right
if abs(result) > 1e12 or math.isnan(result) or math.isinf(result):
return None
formatted_result = format_number(result)
except (OverflowError, ValueError):
return None
part = f"{left_str} {op} {right_str}"
steps.append(random.choice(COT_STEP_TEMPLATES["add_sub"]).format(part=part, result=formatted_result))
current_expr = current_expr[:as_match.start()] + ' ' + formatted_result + ' ' + current_expr[as_match.end():]
reduction_made = True
continue
# If no reduction was made, we're stuck - return None
if not reduction_made:
return None
# Timeout reached
return None
def generate_training_example(_=None):
"""Generate a single training example with retry logic."""
max_retries = 50 # Reduced from 100 for faster generation
for attempt in range(max_retries):
try:
expression_parts = generate_expression_parts()
expression_str = " ".join(map(str, expression_parts))
cot_result = solve_with_cot(expression_str)
if cot_result and isinstance(cot_result['result'], (int, float)):
final_result = cot_result['result']
# Filter out extreme values
if abs(final_result) > 1e12 or (final_result != 0 and abs(final_result) < 1e-4):
continue
if math.isnan(final_result) or math.isinf(final_result):
continue
result_str = format_number(final_result)
if len(result_str) > 20:
continue
use_words = random.random() < WORD_FORM_CHANCE
if use_words:
expression_text = ' '.join([number_to_words(p) if isinstance(p, int) else operator_to_word(p) if isinstance(p, str) else str(p) for p in expression_parts])
result_text = number_to_words(int(round(final_result)))
completion = random.choice(SIMPLE_COMPLETION_TEMPLATES).format(expression=expression_text, result=result_text)
else:
expression_text = expression_str
result_text = result_str
use_reasoning = random.random() < REASONING_CHANCE
if use_reasoning:
intro = random.choice(COT_INTRO_TEMPLATES).format(expression=expression_text)
steps_text = " ".join(cot_result['steps'])
finalizer = random.choice(COT_FINALIZER_TEMPLATES).format(result=result_text)
completion = f"{intro} {steps_text} {finalizer} </think>"
else:
completion = random.choice(SIMPLE_COMPLETION_TEMPLATES).format(expression=expression_text, result=result_text)
if random.random() < SENTENCE_FORM_CHANCE:
prompt = random.choice(PROMPT_TEMPLATES).format(expression=expression_text)
else:
prompt = f"{expression_text} ="
# Clean up spacing
prompt = re.sub(r'\s*\(', ' (', prompt)
prompt = re.sub(r'\)\s*', ') ', prompt).strip()
prompt = re.sub(r'\s+', ' ', prompt)
completion = re.sub(r'\s*\(', ' (', completion)
completion = re.sub(r'\)\s*', ') ', completion).strip()
completion = re.sub(r'\s+', ' ', completion)
return {"prompt": prompt, "completion": " " + completion}
except Exception as e:
continue
return None
def main():
print(f"🔥 Generating {NUM_LINES:,} examples using {NUM_WORKERS} parallel workers...")
print(f" Appending to '{OUTPUT_FILE}'...")
start_time = time.time()
generated_count = 0
failed_count = 0
with open(OUTPUT_FILE, "a", encoding="utf-8") as f:
with multiprocessing.Pool(processes=NUM_WORKERS) as pool:
results_iterator = pool.imap_unordered(generate_training_example, range(NUM_LINES), chunksize=100)
for item in tqdm(results_iterator, total=NUM_LINES, desc="Generating examples"):
if item:
f.write(json.dumps(item) + "\n")
generated_count += 1
else:
failed_count += 1
elapsed_time = time.time() - start_time
print(f"\n\n✅ Done! Appended {generated_count:,} new items to '{OUTPUT_FILE}' in {elapsed_time:.2f}s.")
print(f" 📊 Success rate: {generated_count}/{NUM_LINES} ({100*generated_count/NUM_LINES:.1f}%)")
if failed_count > 0:
print(f" ⚠️ {failed_count:,} generation attempts failed (expressions too complex or invalid)")
if __name__ == "__main__":
multiprocessing.freeze_support()
main() |