int vs float in Python

# No overflow — Python ints grow automatically
big = 2 ** 128
print(big)              # 340282366920938463463374607431768211456
print(type(big))        # <class 'int'>

# Common int operations
a, b = 17, 5
print(a + b)    # 22
print(a - b)    # 12
print(a * b)    # 85
print(a // b)   # 3   — floor division (integer result)
print(a % b)    # 2   — modulo (remainder)
print(a ** b)   # 1419857 — exponentiation
print(a / b)    # 3.4 — always returns float, even if evenly divisible
print(10 / 2)   # 5.0 — float!

x = 3.14
print(type(x))   # <class 'float'>

# Scientific notation
y = 1.5e-3   # 0.0015
z = 2.5e10   # 25000000000.0

# Float operations
print(0.1 + 0.2)          # 0.30000000000000004 — NOT 0.3!
print(0.1 + 0.2 == 0.3)   # False — floating-point representation error

# Correct comparison: use math.isclose
import math
print(math.isclose(0.1 + 0.2, 0.3))           # True
print(math.isclose(0.1 + 0.2, 0.3, rel_tol=1e-9))   # True

# 0.1 in binary is a repeating fraction — it gets rounded
from decimal import Decimal
print(Decimal(0.1))
# 0.1000000000000000055511151231257827021181583404541015625

# The stored value is the closest representable binary fraction to 0.1

# More examples
print(1.1 + 2.2)         # 3.3000000000000003
print(0.1 * 3)           # 0.30000000000000004
print(round(0.1 * 3, 1)) # 0.3 — round() helps for display

# int → float
x = float(5)
print(x)           # 5.0
print(type(x))     # <class 'float'>

# float → int: truncates (does NOT round)
print(int(3.9))    # 3 — truncated, not rounded
print(int(-3.9))   # -3 — truncated toward zero
print(int(3.0))    # 3

# Rounding
print(round(3.9))        # 4
print(round(3.14159, 2)) # 3.14

# Banker's rounding (rounds to nearest even)
print(round(2.5))   # 2 — rounds to even
print(round(3.5))   # 4 — rounds to even

import math

# Infinity
pos_inf = float("inf")
neg_inf = float("-inf")
print(pos_inf > 1e308)   # True
print(neg_inf < -1e308)  # True

# Not a Number
nan = float("nan")
print(math.isnan(nan))   # True
print(nan == nan)         # False — NaN is not equal to itself!
print(math.isnan(float("nan")))  # True — correct way to check

# math constants
print(math.pi)    # 3.141592653589793
print(math.e)     # 2.718281828459045
print(math.inf)   # inf
print(math.nan)   # nan

import math

def is_loss_improving(old_loss: float, new_loss: float, min_delta: float = 1e-6) -> bool:
    """Check if loss improved by more than a threshold."""
    return (old_loss - new_loss) > min_delta   # Never use == for floats

def is_valid_loss(loss: float) -> bool:
    """A NaN or infinite loss signals a training crash."""
    return math.isfinite(loss)

# Training loop guard
loss = compute_loss(model, batch)
if not is_valid_loss(loss):
    raise RuntimeError(f"Training diverged: loss = {loss}")

def clamp_probability(p: float) -> float:
    """Clamp to valid probability range [0, 1]."""
    return max(0.0, min(1.0, p))

def log_loss(y_true: int, y_pred: float, epsilon: float = 1e-7) -> float:
    """
    Binary cross-entropy loss.
    epsilon guards against log(0) which is -inf.
    """
    import math
    y_pred = clamp_probability(y_pred)
    y_pred = max(epsilon, min(1 - epsilon, y_pred))  # Clip away from 0 and 1
    return -(y_true * math.log(y_pred) + (1 - y_true) * math.log(1 - y_pred))

print(log_loss(1, 0.9))   # 0.10536... — low loss, correct prediction
print(log_loss(1, 0.1))   # 2.30258... — high loss, wrong prediction

import numpy as np

embeddings = np.random.randn(100, 1536)

# int: fine for indexing
idx = 5
print(embeddings[idx].shape)   # (1536,)

# float: NOT valid as an index
ratio = 0.8
split = int(len(embeddings) * ratio)   # Convert to int first
train = embeddings[:split]
test  = embeddings[split:]

print(train.shape)   # (80, 1536)
print(test.shape)    # (20, 1536)

def estimate_tokens(text: str) -> int:
    """Rough estimate: ~4 characters per token. Always returns int."""
    return len(text) // 4   # // ensures int result

def within_budget(text: str, max_tokens: int = 4096) -> bool:
    return estimate_tokens(text) <= max_tokens

# Integer division is exact and safe for this
budget_ratio = 0.8
max_context = 8192
soft_limit = int(max_context * budget_ratio)   # 6553 — always convert
print(soft_limit)   # 6553

# Pitfall 1: using == to compare floats
total = 0.1 + 0.1 + 0.1
if total == 0.3:              # False! Don't do this
    print("equal")
if math.isclose(total, 0.3):  # True — do this instead
    print("approximately equal")


# Pitfall 2: int division gives float in Python 3
result = 10 / 3
print(result, type(result))   # 3.3333... <class 'float'>

# Use // for integer (floor) division
result = 10 // 3
print(result, type(result))   # 3 <class 'int'>


# Pitfall 3: NaN silently propagates
import numpy as np
scores = np.array([0.9, float("nan"), 0.8])
print(scores.mean())       # nan — the whole mean is poisoned
print(np.nanmean(scores))  # 0.85 — ignores NaN


# Pitfall 4: loss of precision in long float chains
total = 0.0
for _ in range(10_000):
    total += 0.1
print(total)               # 1000.0000000000159 — slight drift
print(math.isclose(total, 1000.0, rel_tol=1e-9))  # False