agillm-math-numeracy-synthetic / generate_math_numeracy.py
OpenTransformer's picture
Add synthetic numeracy pretraining corpus
6d600ca verified
#!/usr/bin/env python3
import json, random, hashlib, time
from pathlib import Path
OUT = Path('/workspace/agillm_math_numeracy_synth/train.jsonl')
META = Path('/workspace/agillm_math_numeracy_synth/metadata.json')
README = Path('/workspace/agillm_math_numeracy_synth/README.md')
N = 2_000_000
SEED = 731095
rng = random.Random(SEED)
objects = ['apples','coins','books','marbles','pencils','blocks','tokens','stickers']
names = ['Ada','Ben','Cara','Dion','Eli','Fay','Gus','Hana']
counts = {}
def jwrite(f, text, kind):
counts[kind] = counts.get(kind, 0) + 1
f.write(json.dumps({'text': text.strip() + '\n', 'kind': kind}, ensure_ascii=True, separators=(',', ':')) + '\n')
def choose(xs): return xs[rng.randrange(len(xs))]
def gen_add(big=False):
m = 9999 if big else 100
a,b = rng.randint(0,m), rng.randint(0,m); c=a+b
return choose([f'{a} + {b} = {c}.', f'{a}+{b}={c}', f'Compute {a} + {b}. Answer: {c}.', f'What is {a} plus {b}? {c}.', f'The sum of {a} and {b} is {c}.']), 'multi_digit_addition' if big else 'addition'
def gen_sub():
b,c = rng.randint(0,1000), rng.randint(0,1000); a=b+c
return choose([f'{a} - {b} = {c}.', f'{a}-{b}={c}', f'Compute {a} minus {b}. Answer: {c}.', f'What is {a} - {b}? {c}.']), 'subtraction'
def gen_mul():
a,b = (rng.randint(0,20), rng.randint(0,20)) if rng.random()<0.75 else (rng.randint(0,99), rng.randint(0,12))
c=a*b
return choose([f'{a} * {b} = {c}.', f'{a} times {b} equals {c}.', f'Compute {a} x {b}. Answer: {c}.', f'The product of {a} and {b} is {c}.']), 'multiplication'
def gen_div():
b,c = rng.randint(1,20), rng.randint(0,100); a=b*c
return choose([f'{a} / {b} = {c}.', f'{a} divided by {b} equals {c}.', f'Compute {a} / {b}. Answer: {c}.']), 'exact_division'
def gen_compare():
a,b = rng.randint(-200,200), rng.randint(-200,200)
sym = '>' if a>b else '<' if a<b else '='
word = 'greater than' if a>b else 'less than' if a<b else 'equal to'
return choose([f'{a} {sym} {b}.', f'{a} is {word} {b}.', f'Compare {a} and {b}: {a} {sym} {b}.']), 'comparison'
def gen_order_parity():
n = rng.randint(-1000,1000)
if rng.random()<0.5:
return choose([f'The number after {n} is {n+1}.', f'The number before {n} is {n-1}.', f'{n-1}, {n}, {n+1}.']), 'number_order'
par = 'even' if n % 2 == 0 else 'odd'
return f'{n} is {par}.', 'parity'
def gen_place():
n = rng.randint(0,9999); ones=n%10; tens=(n//10)%10; hundreds=(n//100)%10; thousands=(n//1000)%10
return choose([f'In {n}, the ones digit is {ones}.', f'In {n}, the tens digit is {tens}.', f'In {n}, the hundreds digit is {hundreds}.', f'In {n}, the thousands digit is {thousands}.', f'{n} = {thousands} thousands + {hundreds} hundreds + {tens} tens + {ones} ones.']), 'place_value'
def gen_missing():
a = rng.randint(0,100); x = rng.randint(0,100); b = a+x
if rng.random()<0.5:
return f'Solve x + {a} = {b}. x = {x}. Check: {x} + {a} = {b}.', 'missing_number'
return f'Solve {b} - x = {a}. x = {x}. Check: {b} - {x} = {a}.', 'missing_number'
def gen_word():
name=choose(names); obj=choose(objects); a=rng.randint(0,30); b=rng.randint(0,30)
if rng.random()<0.55:
return f'{name} has {a} {obj} and gets {b} more {obj}. {name} has {a+b} {obj}.', 'word_addition'
if b>a: a,b=b,a
return f'{name} has {a} {obj} and gives away {b}. {name} has {a-b} {obj} left.', 'word_subtraction'
def gen_multistep():
a,b,c = rng.randint(0,50), rng.randint(0,50), rng.randint(0,50)
if rng.random()<0.5:
r=a+b-c
return f'{a} + {b} - {c} = {r}. First {a} + {b} = {a+b}; then {a+b} - {c} = {r}.', 'two_step'
r=a*b+c
return f'{a} * {b} + {c} = {r}. First {a} * {b} = {a*b}; then {a*b} + {c} = {r}.', 'two_step'
def gen_fraction():
d=rng.randint(2,12); n=rng.randint(1,d-1); k=rng.randint(2,10)
return f'{n}/{d} = {n*k}/{d*k}. Multiplying numerator and denominator by {k} makes an equivalent fraction.', 'fractions'
def gen_money():
cents=rng.randint(0,9999); dollars=cents//100; rem=cents%100
return f'{cents} cents = {dollars} dollars and {rem} cents.', 'money'
gens = [gen_add, lambda: gen_add(True), gen_sub, gen_mul, gen_div, gen_compare, gen_order_parity, gen_place, gen_missing, gen_word, gen_multistep, gen_fraction, gen_money]
weights = [0.18,0.07,0.14,0.14,0.08,0.08,0.08,0.06,0.07,0.07,0.05,0.04,0.04]
def weighted_gen():
total = sum(weights)
r=rng.random() * total; acc=0.0
for g,w in zip(gens,weights):
acc += w
if r <= acc: return g()
return gens[-1]()
start=time.time(); sha=hashlib.sha256(); rows=0
with OUT.open('w', encoding='utf-8') as f:
# Exhaustive core facts first, repeated in clean forms.
for a in range(0,101):
for b in range(0,101):
jwrite(f, f'{a} + {b} = {a+b}.', 'addition_table'); rows+=1
if a >= b:
jwrite(f, f'{a} - {b} = {a-b}.', 'subtraction_table'); rows+=1
for a in range(0,21):
for b in range(0,21):
jwrite(f, f'{a} * {b} = {a*b}.', 'multiplication_table'); rows+=1
if b != 0:
jwrite(f, f'{a*b} / {b} = {a}.', 'division_table'); rows+=1
while rows < N:
text, kind = weighted_gen()
jwrite(f, text, kind); rows += 1
if rows % 250000 == 0:
print(f'generated {rows:,}', flush=True)
# Hash after close.
with OUT.open('rb') as f:
for chunk in iter(lambda: f.read(1<<20), b''):
sha.update(chunk)
meta = {'name':'agillm-math-numeracy-synthetic','rows':rows,'seed':SEED,'sha256':sha.hexdigest(),'counts':counts,'created_unix':int(time.time()),'description':'Synthetic numeracy pretraining JSONL with exact arithmetic facts, comparisons, place value, word problems, and simple step-by-step arithmetic.'}
META.write_text(json.dumps(meta, indent=2, sort_keys=True) + '\n')
README.write_text('''---\nlicense: apache-2.0\ntags:\n- synthetic\n- math\n- numeracy\n- pretraining\n---\n\n# AGILLM Math Numeracy Synthetic\n\nSynthetic plain-text numeracy corpus for AGILLM pretraining. Each JSONL row has `text` and `kind`.\n\nCoverage includes addition/subtraction/multiplication/division facts, multi-digit arithmetic, comparisons, order/parity, place value, missing-number equations, simple word problems, fractions, money, and two-step arithmetic.\n\nGenerated deterministically with `generate_math_numeracy.py`.\n''')
print(json.dumps(meta, indent=2, sort_keys=True))
print(f'elapsed_sec={time.time()-start:.1f}')