Datasets:

leejason2025
/

magic-cap1b-data

from __future__ import annotations
import re

# ---------------------------------------------------------------------------
# Vocabulary
# ---------------------------------------------------------------------------
# Base: 256 raw bytes. Token id is identical to the byte value, so
# encode/decode of arbitrary text is trivial and lossless.
BYTE_TOKENS = [f"<BYTE_{i}>" for i in range(256)]
N_BYTES = 256
 
# Reserved atomic tokens, ids 256+
DIGIT_TOKENS = [f"<DIGIT_{d}>" for d in range(10)]
OP_TOKENS = ["<ADD>", "<SUB>", "<MULTIPLY>"]
MARKER_TOKENS = ["<START>", "<STEP>", "<FINAL>", "<PIPE>", "<COLON>"]
STRUCTURAL_TOKENS = ["<BOUNDARY>", "<PAD>"]
 
RESERVED_TOKENS = DIGIT_TOKENS + OP_TOKENS + MARKER_TOKENS + STRUCTURAL_TOKENS
ALL_TOKENS = BYTE_TOKENS + RESERVED_TOKENS
 
TOKEN_TO_ID: dict[str, int] = {t: i for i, t in enumerate(ALL_TOKENS)}
ID_TO_TOKEN: dict[int, str] = {i: t for t, i in TOKEN_TO_ID.items()}
VOCAB_SIZE = len(ALL_TOKENS)          # 276
 
DIGIT_BASE = TOKEN_TO_ID["<DIGIT_0>"]  # 256
BOUNDARY_ID = TOKEN_TO_ID["<BOUNDARY>"]
PAD_ID = TOKEN_TO_ID["<PAD>"]
 
# Reserved-token id range, used by the invariant checks.
RESERVED_ID_MIN = N_BYTES
RESERVED_ID_MAX = VOCAB_SIZE - 1
 
# Canonical operation name -> reserved token. Aliases are normalised so the
# tokenizer does not care which spelling generator.py emits.
_OP_ALIASES = {
    "ADD": "<ADD>",
    "SUB": "<SUB>",
    "SUBTRACT": "<SUB>",
    "MUL": "<MULTIPLY>",
    "MULT": "<MULTIPLY>",
    "MULTIPLY": "<MULTIPLY>",
}
 
# Surface strings used when decoding reserved tokens back to text.
_RESERVED_SURFACE = {
    "<ADD>": "ADD",
    "<SUB>": "SUB",
    "<MULTIPLY>": "MULTIPLY",
    "<START>": "START",
    "<STEP>": "STEP",
    "<FINAL>": "FINAL",
    "<PIPE>": "|",
    "<COLON>": ":",
    "<BOUNDARY>": "\n",
    "<PAD>": "",
}
 
 
# ---------------------------------------------------------------------------
# Arithmetic-safety validation  (B.4 exclusions)
# ---------------------------------------------------------------------------
_BANNED_PATTERNS = [
    (re.compile(r"(?<=\d),(?=\d)"), "comma inside a number"),
    (re.compile(r"(?<=\d)\.(?=\d)"), "decimal point inside a number"),
]
 
# A minus is UNARY (banned) only when nothing operand-like precedes it. After a
# variable, a digit, or a closing paren it is BINARY subtraction, which the code
# family legitimately emits as `x = x - 9`.
_MINUS_BEFORE_DIGIT = re.compile(r"-\s*\d")
_OPERAND_END = re.compile(r"[\w)\]]")
 
 
def assert_arithmetic_safe(text: str, where: str = "") -> None:
    """Raise if text violates a B.4 exclusion. Cheap, run on every record."""
    for pattern, reason in _BANNED_PATTERNS:
        if pattern.search(text):
            raise ValueError(f"B.4 violation ({reason}) in {where or 'text'}: {text!r}")
 
    for m in _MINUS_BEFORE_DIGIT.finditer(text):
        prefix = text[:m.start()].rstrip()
        if prefix and _OPERAND_END.match(prefix[-1]):
            continue  # binary subtraction
        raise ValueError(
            f"B.4 violation (unary negative sign before a number) in "
            f"{where or 'text'}: {text!r}"
        )
 
 
# ---------------------------------------------------------------------------
# Core encoders
# ---------------------------------------------------------------------------
def encode_number(n: int | str) -> list[int]:
    """
    Canonical digit-by-digit serialization. `47` -> [<DIGIT_4>, <DIGIT_7>].
    Reads the value it is given; performs no arithmetic.
    """
    s = str(n)
    if not s.isdigit():
        raise ValueError(f"non-canonical number {n!r}: must be a nonnegative integer")
    return [DIGIT_BASE + int(ch) for ch in s]
 
 
def encode_text(text: str) -> list[int]:
    """
    General byte-level path with the digit override applied.
 
    Every byte of the UTF-8 encoding becomes one token, except ASCII digits,
    which become atomic <DIGIT_n> tokens. Byte ids 48-57 are therefore never
    emitted, which is what `assert_digits_atomic` checks.
 
    Emits no reserved marker or operator tokens. Handles arbitrary text,
    including FineWeb-Edu background.
    """
    out: list[int] = []
    for b in text.encode("utf-8"):
        if 48 <= b <= 57:                 # ASCII '0'-'9'
            out.append(DIGIT_BASE + (b - 48))
        else:
            out.append(b)
    return out
 
 
# ---------------------------------------------------------------------------
# Canonical specification path  (B.2 format, atomic markers)
#   START: 17 | STEP_1: ADD 6 | STEP_2: MULTIPLY 2 | FINAL: 46
# ---------------------------------------------------------------------------
_SEG_START = re.compile(r"^START\s*:\s*(\d+)$")
_SEG_STEP = re.compile(r"^STEP_(\d+)\s*:\s*([A-Za-z_]+)\s+(\d+)$")
_SEG_FINAL = re.compile(r"^FINAL\s*:\s*(\d*)$")
 
 
def parse_canonical(text: str) -> dict:
    """
    Parse a canonical spec string into its stored parts. Pure string reading,
    no arithmetic, so this respects the renderer-never-computes contract.
 
    Returns {"start": str, "steps": [(index, OP, operand)], "final": str|None}
    """
    start = None
    steps: list[tuple[str, str, str]] = []
    final = None
 
    for raw in text.replace("\n", " ").split("|"):
        seg = raw.strip()
        if not seg:
            continue
        m = _SEG_START.match(seg)
        if m:
            start = m.group(1)
            continue
        m = _SEG_STEP.match(seg)
        if m:
            op = m.group(2).upper()
            if op not in _OP_ALIASES:
                raise ValueError(f"unknown operation {op!r} in canonical spec: {text!r}")
            steps.append((m.group(1), op, m.group(3)))
            continue
        m = _SEG_FINAL.match(seg)
        if m:
            final = m.group(1) or None
            continue
        raise ValueError(f"unparseable canonical segment {seg!r} in: {text!r}")
 
    if start is None:
        raise ValueError(f"canonical spec has no START: {text!r}")
    return {"start": start, "steps": steps, "final": final}
 
 
def encode_canonical(text: str) -> list[int]:
    """Encode a canonical spec string with atomic markers, operators, digits."""
    return encode_canonical_parts(parse_canonical(text))
 
 
def encode_canonical_parts(parts: dict) -> list[int]:
    ids: list[int] = [TOKEN_TO_ID["<START>"], TOKEN_TO_ID["<COLON>"]]
    ids += encode_number(parts["start"])
    for idx, op, operand in parts["steps"]:
        ids.append(TOKEN_TO_ID["<PIPE>"])
        ids.append(TOKEN_TO_ID["<STEP>"])
        ids += encode_number(idx)
        ids.append(TOKEN_TO_ID["<COLON>"])
        ids.append(TOKEN_TO_ID[_OP_ALIASES[op]])
        ids += encode_number(operand)
    ids.append(TOKEN_TO_ID["<PIPE>"])
    ids.append(TOKEN_TO_ID["<FINAL>"])
    ids.append(TOKEN_TO_ID["<COLON>"])
    if parts["final"] is not None:
        ids += encode_number(parts["final"])
    return ids
 
 
# ---------------------------------------------------------------------------
# Record routing
# ---------------------------------------------------------------------------
def encode_record(record: dict) -> list[int]:
    """
    Encode one generator.py record.
 
    Routing, and the reason for it: the `direct` family's surface text IS the
    canonical specification, and B.4 wants atomic operator and marker tokens
    there. Every other family has its own distinct surface text (verbal prose,
    code trace, decomposed primitives) and goes through the byte-level path
    on its OWN text.
 
    Routing every family through the canonical encoder would collapse all
    four families to near-identical token streams and destroy the family
    contrast the whole experiment measures. `test_families_tokenize_differently`
    is the regression guard for that.
    """
    family = record["family"]
    text = record["text"]
    assert_arithmetic_safe(text, where=f"{family}/{record.get('program_id')}")
    if family == "direct":
        return encode_canonical(text)
    return encode_text(text)
 
 
# ---------------------------------------------------------------------------
# Decoding and invariants
# ---------------------------------------------------------------------------
def decode(ids: list[int]) -> str:
    """Inverse of encode_text; reserved tokens decode to their surface form."""
    out: list[str] = []
    buf = bytearray()
 
    def flush() -> None:
        if buf:
            out.append(buf.decode("utf-8", errors="replace"))
            buf.clear()
 
    for i in ids:
        if i < N_BYTES:
            buf.append(i)
            continue
        flush()
        tok = ID_TO_TOKEN[i]
        if tok.startswith("<DIGIT_"):
            out.append(tok[len("<DIGIT_"):-1])
        else:
            out.append(_RESERVED_SURFACE[tok])
    flush()
    return "".join(out)
 
 
def assert_digits_atomic(ids: list[int]) -> None:
    """
    Real check, not a tautology: a digit character may only ever appear as a
    <DIGIT_n> token, so raw byte ids 48-57 must never occur in output.
    """
    bad = [i for i in ids if 48 <= i <= 57]
    if bad:
        raise AssertionError(f"raw digit byte ids leaked into output: {sorted(set(bad))}")
 
 
def emits_reserved_markers(ids: list[int]) -> bool:
    """True if the sequence contains any operator or task-marker token.
    Used by the contamination check: only synthetic data can produce these."""
    marker_ids = {TOKEN_TO_ID[t] for t in OP_TOKENS + MARKER_TOKENS}
    return any(i in marker_ids for i in ids)

	from __future__ import annotations
	import re

	# ---------------------------------------------------------------------------
	# Vocabulary
	# ---------------------------------------------------------------------------
	# Base: 256 raw bytes. Token id is identical to the byte value, so
	# encode/decode of arbitrary text is trivial and lossless.
	BYTE_TOKENS = [f"<BYTE_{i}>" for i in range(256)]
	N_BYTES = 256

	# Reserved atomic tokens, ids 256+
	DIGIT_TOKENS = [f"<DIGIT_{d}>" for d in range(10)]
	OP_TOKENS = ["<ADD>", "<SUB>", "<MULTIPLY>"]
	MARKER_TOKENS = ["<START>", "<STEP>", "<FINAL>", "<PIPE>", "<COLON>"]
	STRUCTURAL_TOKENS = ["<BOUNDARY>", "<PAD>"]

	RESERVED_TOKENS = DIGIT_TOKENS + OP_TOKENS + MARKER_TOKENS + STRUCTURAL_TOKENS
	ALL_TOKENS = BYTE_TOKENS + RESERVED_TOKENS

	TOKEN_TO_ID: dict[str, int] = {t: i for i, t in enumerate(ALL_TOKENS)}
	ID_TO_TOKEN: dict[int, str] = {i: t for t, i in TOKEN_TO_ID.items()}
	VOCAB_SIZE = len(ALL_TOKENS) # 276

	DIGIT_BASE = TOKEN_TO_ID["<DIGIT_0>"] # 256
	BOUNDARY_ID = TOKEN_TO_ID["<BOUNDARY>"]
	PAD_ID = TOKEN_TO_ID["<PAD>"]

	# Reserved-token id range, used by the invariant checks.
	RESERVED_ID_MIN = N_BYTES
	RESERVED_ID_MAX = VOCAB_SIZE - 1

	# Canonical operation name -> reserved token. Aliases are normalised so the
	# tokenizer does not care which spelling generator.py emits.
	_OP_ALIASES = {
	"ADD": "<ADD>",
	"SUB": "<SUB>",
	"SUBTRACT": "<SUB>",
	"MUL": "<MULTIPLY>",
	"MULT": "<MULTIPLY>",
	"MULTIPLY": "<MULTIPLY>",
	}

	# Surface strings used when decoding reserved tokens back to text.
	_RESERVED_SURFACE = {
	"<ADD>": "ADD",
	"<SUB>": "SUB",
	"<MULTIPLY>": "MULTIPLY",
	"<START>": "START",
	"<STEP>": "STEP",
	"<FINAL>": "FINAL",
	"<PIPE>": "\|",
	"<COLON>": ":",
	"<BOUNDARY>": "\n",
	"<PAD>": "",
	}


	# ---------------------------------------------------------------------------
	# Arithmetic-safety validation (B.4 exclusions)
	# ---------------------------------------------------------------------------
	_BANNED_PATTERNS = [
	(re.compile(r"(?<=\d),(?=\d)"), "comma inside a number"),
	(re.compile(r"(?<=\d)\.(?=\d)"), "decimal point inside a number"),
	]

	# A minus is UNARY (banned) only when nothing operand-like precedes it. After a
	# variable, a digit, or a closing paren it is BINARY subtraction, which the code
	# family legitimately emits as `x = x - 9`.
	_MINUS_BEFORE_DIGIT = re.compile(r"-\s*\d")
	_OPERAND_END = re.compile(r"[\w)\]]")


	def assert_arithmetic_safe(text: str, where: str = "") -> None:
	"""Raise if text violates a B.4 exclusion. Cheap, run on every record."""
	for pattern, reason in _BANNED_PATTERNS:
	if pattern.search(text):
	raise ValueError(f"B.4 violation ({reason}) in {where or 'text'}: {text!r}")

	for m in _MINUS_BEFORE_DIGIT.finditer(text):
	prefix = text[:m.start()].rstrip()
	if prefix and _OPERAND_END.match(prefix[-1]):
	continue # binary subtraction
	raise ValueError(
	f"B.4 violation (unary negative sign before a number) in "
	f"{where or 'text'}: {text!r}"
	)


	# ---------------------------------------------------------------------------
	# Core encoders
	# ---------------------------------------------------------------------------
	def encode_number(n: int \| str) -> list[int]:
	"""
	Canonical digit-by-digit serialization. `47` -> [<DIGIT_4>, <DIGIT_7>].
	Reads the value it is given; performs no arithmetic.
	"""
	s = str(n)
	if not s.isdigit():
	raise ValueError(f"non-canonical number {n!r}: must be a nonnegative integer")
	return [DIGIT_BASE + int(ch) for ch in s]


	def encode_text(text: str) -> list[int]:
	"""
	General byte-level path with the digit override applied.

	Every byte of the UTF-8 encoding becomes one token, except ASCII digits,
	which become atomic <DIGIT_n> tokens. Byte ids 48-57 are therefore never
	emitted, which is what `assert_digits_atomic` checks.

	Emits no reserved marker or operator tokens. Handles arbitrary text,
	including FineWeb-Edu background.
	"""
	out: list[int] = []
	for b in text.encode("utf-8"):
	if 48 <= b <= 57: # ASCII '0'-'9'
	out.append(DIGIT_BASE + (b - 48))
	else:
	out.append(b)
	return out


	# ---------------------------------------------------------------------------
	# Canonical specification path (B.2 format, atomic markers)
	# START: 17 \| STEP_1: ADD 6 \| STEP_2: MULTIPLY 2 \| FINAL: 46
	# ---------------------------------------------------------------------------
	_SEG_START = re.compile(r"^START\s:\s(\d+)$")
	_SEG_STEP = re.compile(r"^STEP_(\d+)\s:\s([A-Za-z_]+)\s+(\d+)$")
	_SEG_FINAL = re.compile(r"^FINAL\s:\s(\d*)$")


	def parse_canonical(text: str) -> dict:
	"""
	Parse a canonical spec string into its stored parts. Pure string reading,
	no arithmetic, so this respects the renderer-never-computes contract.

	Returns {"start": str, "steps": [(index, OP, operand)], "final": str\|None}
	"""
	start = None
	steps: list[tuple[str, str, str]] = []
	final = None

	for raw in text.replace("\n", " ").split("\|"):
	seg = raw.strip()
	if not seg:
	continue
	m = _SEG_START.match(seg)
	if m:
	start = m.group(1)
	continue
	m = _SEG_STEP.match(seg)
	if m:
	op = m.group(2).upper()
	if op not in _OP_ALIASES:
	raise ValueError(f"unknown operation {op!r} in canonical spec: {text!r}")
	steps.append((m.group(1), op, m.group(3)))
	continue
	m = _SEG_FINAL.match(seg)
	if m:
	final = m.group(1) or None
	continue
	raise ValueError(f"unparseable canonical segment {seg!r} in: {text!r}")

	if start is None:
	raise ValueError(f"canonical spec has no START: {text!r}")
	return {"start": start, "steps": steps, "final": final}


	def encode_canonical(text: str) -> list[int]:
	"""Encode a canonical spec string with atomic markers, operators, digits."""
	return encode_canonical_parts(parse_canonical(text))


	def encode_canonical_parts(parts: dict) -> list[int]:
	ids: list[int] = [TOKEN_TO_ID["<START>"], TOKEN_TO_ID["<COLON>"]]
	ids += encode_number(parts["start"])
	for idx, op, operand in parts["steps"]:
	ids.append(TOKEN_TO_ID["<PIPE>"])
	ids.append(TOKEN_TO_ID["<STEP>"])
	ids += encode_number(idx)
	ids.append(TOKEN_TO_ID["<COLON>"])
	ids.append(TOKEN_TO_ID[_OP_ALIASES[op]])
	ids += encode_number(operand)
	ids.append(TOKEN_TO_ID["<PIPE>"])
	ids.append(TOKEN_TO_ID["<FINAL>"])
	ids.append(TOKEN_TO_ID["<COLON>"])
	if parts["final"] is not None:
	ids += encode_number(parts["final"])
	return ids


	# ---------------------------------------------------------------------------
	# Record routing
	# ---------------------------------------------------------------------------
	def encode_record(record: dict) -> list[int]:
	"""
	Encode one generator.py record.

	Routing, and the reason for it: the `direct` family's surface text IS the
	canonical specification, and B.4 wants atomic operator and marker tokens
	there. Every other family has its own distinct surface text (verbal prose,
	code trace, decomposed primitives) and goes through the byte-level path
	on its OWN text.

	Routing every family through the canonical encoder would collapse all
	four families to near-identical token streams and destroy the family
	contrast the whole experiment measures. `test_families_tokenize_differently`
	is the regression guard for that.
	"""
	family = record["family"]
	text = record["text"]
	assert_arithmetic_safe(text, where=f"{family}/{record.get('program_id')}")
	if family == "direct":
	return encode_canonical(text)
	return encode_text(text)


	# ---------------------------------------------------------------------------
	# Decoding and invariants
	# ---------------------------------------------------------------------------
	def decode(ids: list[int]) -> str:
	"""Inverse of encode_text; reserved tokens decode to their surface form."""
	out: list[str] = []
	buf = bytearray()

	def flush() -> None:
	if buf:
	out.append(buf.decode("utf-8", errors="replace"))
	buf.clear()

	for i in ids:
	if i < N_BYTES:
	buf.append(i)
	continue
	flush()
	tok = ID_TO_TOKEN[i]
	if tok.startswith("<DIGIT_"):
	out.append(tok[len("<DIGIT_"):-1])
	else:
	out.append(_RESERVED_SURFACE[tok])
	flush()
	return "".join(out)


	def assert_digits_atomic(ids: list[int]) -> None:
	"""
	Real check, not a tautology: a digit character may only ever appear as a
	<DIGIT_n> token, so raw byte ids 48-57 must never occur in output.
	"""
	bad = [i for i in ids if 48 <= i <= 57]
	if bad:
	raise AssertionError(f"raw digit byte ids leaked into output: {sorted(set(bad))}")


	def emits_reserved_markers(ids: list[int]) -> bool:
	"""True if the sequence contains any operator or task-marker token.
	Used by the contamination check: only synthetic data can produce these."""
	marker_ids = {TOKEN_TO_ID[t] for t in OP_TOKENS + MARKER_TOKENS}
	return any(i in marker_ids for i in ids)