Stage B/C datasets v1.0.0: 25k deep-CoT + 10k exec problems (+verify twin), manifests, audit, generators

cafb492 verified 25 days ago

13 kB

"""AGA-Code Stage B/C dataset framework.
Everything here is original, procedurally generated, and execution-verified.
No distillation, no third-party traces — correctness is established by
RUNNING the code, never by copying anyone's answer.
Execution harness semantics are byte-identical to aga_code.run_tests:
    harness = code + "\\n" + tests + "\\nprint('AGA_EXEC_OK')"
    pass = returncode == 0 and sentinel in stdout
"""
from __future__ import annotations

import hashlib
import random
import subprocess
import sys
from dataclasses import dataclass, field

MASTER_SEED = 1147
SENTINEL = "AGA_EXEC_OK"


# --------------------------------------------------------------------------
# Execution harness (identical semantics to aga_code.run_tests)
# --------------------------------------------------------------------------
def run_code(code: str, tests: str, timeout: int = 5):
    """Run code+tests in a subprocess. Returns (passed: bool, output: str)."""
    harness = code + "\n" + tests + "\nprint('AGA_EXEC_OK')"
    try:
        r = subprocess.run([sys.executable, "-c", harness],
                           capture_output=True, timeout=timeout, text=True)
        ok = r.returncode == 0 and SENTINEL in (r.stdout or "")
        return ok, (r.stdout or "") + (r.stderr or "")
    except subprocess.TimeoutExpired:
        return False, "TIMEOUT"
    except Exception as e:  # loud, never silent
        return False, f"HARNESS ERROR: {e}"


def run_tests(code: str, tests: str, timeout: int = 5) -> bool:
    return run_code(code, tests, timeout)[0]


class _ExecTimeout(BaseException):
    pass


def _alarm_handler(signum, frame):
    raise _ExecTimeout()


def exec_run(code: str, tests: str, timeout: float = 2.0) -> bool:
    """Fast in-process execution gate with identical pass/fail semantics to
    run_tests (code + tests + sentinel print; pass = completes cleanly).
    Used for the high-volume build gates; the shipped data is additionally
    re-verified with the true subprocess harness (driver spot-check + the
    independent audit). BaseException timeout so mutant code cannot swallow
    it with a bare `except Exception`.
    """
    import signal
    import contextlib
    import io
    src = code + "\n" + tests + "\nprint('AGA_EXEC_OK')"
    try:
        compiled = compile(src, "<gate>", "exec")
    except SyntaxError:
        return False
    old = signal.signal(signal.SIGALRM, _alarm_handler)
    signal.setitimer(signal.ITIMER_REAL, timeout)
    try:
        with contextlib.redirect_stdout(io.StringIO()):
            exec(compiled, {})
        return True
    except BaseException:
        return False
    finally:
        signal.setitimer(signal.ITIMER_REAL, 0)
        signal.signal(signal.SIGALRM, old)


def first_failing_assert(code: str, tests: str, timeout: int = 5):
    """Return the first single test line that fails against `code`, or None.
    Non-assert lines (e.g. fixture setup like `s = Stack()`) are prepended as
    setup context so the failing assertion reflects the real bug.
    """
    setup = [ln for ln in tests.splitlines()
             if ln.strip() and not ln.strip().startswith("assert")]
    asserts = [ln for ln in tests.splitlines() if ln.strip().startswith("assert")]
    ctx = "\n".join(setup)
    for ln in asserts:
        ok, _ = run_code(code, ctx + "\n" + ln if ctx else ln, timeout)
        if not ok:
            return ln.strip()
    return None


# --------------------------------------------------------------------------
# Data structures
# --------------------------------------------------------------------------
@dataclass
class Bug:
    code: str          # buggy variant of the reference
    kind: str          # off_by_one | wrong_operator | wrong_edge | swapped_branch | missing_case
    note: str          # family-provided diagnosis detail (what is actually wrong)


@dataclass
class Gen:
    params: dict
    instruction: str                    # fully rendered instruction text
    phrasing: str                       # which instruction template produced it
    think_slots: dict                   # goal/approach/steps/edges/complexity
    reference: str                      # correct solution code
    tests: str                          # newline-joined assert statements
    difficulty: str                     # easy | medium | hard (per-instance)
    bugs: list = field(default_factory=list)   # list[Bug]
    func: str = ""                      # primary function/class name


@dataclass
class Family:
    name: str
    difficulty: str                     # base: easy | medium | hard
    instruction_templates: list         # list[str] with {slots}
    gen: object                         # callable(rng, hard: bool) -> Gen


REGISTRY: list = []


def family(name, difficulty, instruction_templates):
    def deco(fn):
        REGISTRY.append(Family(name=name, difficulty=difficulty,
                               instruction_templates=instruction_templates, gen=fn))
        return fn
    return deco


# --------------------------------------------------------------------------
# Think skeletons — 12 varied phrasings filled from family-provided slots.
# Slots: {goal} {approach} {steps} {edges} {complexity}
# --------------------------------------------------------------------------
THINK_SKELETONS = [
    "The task is to {goal}. My plan: {approach}. Concretely: {steps}. I have to be careful with {edges}. Overall this runs in {complexity}.",
    "Breaking it down: {goal}. The clean way to do this is {approach}. Steps: {steps}. Watch out for {edges}. Time complexity: {complexity}.",
    "Plan first. Goal: {goal}. Approach: {approach}. Then {steps}. Edge cases that could bite: {edges}. Expected complexity {complexity}.",
    "I need to {goal}. Strategy: {approach}. Walkthrough: {steps}. Before finalizing, check {edges}. This should be {complexity}.",
    "Let me think about {goal}. A direct approach is {approach}: {steps}. The risky parts are {edges}. Complexity is {complexity}.",
    "Decomposition: {goal}. Main idea: {approach}. Sub-steps: {steps}. Handle {edges} explicitly. Cost: {complexity}.",
    "For {goal}, I'll {approach}. The algorithm: {steps}. Non-obvious cases: {edges}. That gives {complexity}.",
    "Thinking it through: to {goal}, the natural method is {approach}. Sequence: {steps}. Don't forget {edges}. {complexity} overall.",
    "Approach for {goal}: {approach}. In order: {steps}. Special attention to {edges}. Complexity: {complexity}.",
    "First principles: {goal}. I'll {approach} because it is straightforward and correct: {steps}. Verify against {edges}. Runs in {complexity}.",
    "Target: {goal}. Method: {approach}. Implementation sketch: {steps}. Failure modes to avoid: {edges}. {complexity}.",
    "Here's my reasoning for {goal}: {approach} is the right tool. Steps: {steps}. Edge review: {edges}. Complexity {complexity}.",
]

# --------------------------------------------------------------------------
# Critique templates per bug kind — filled with the family's diagnosis note
# and the ACTUAL failing assertion observed by execution.
# Bug kinds mirror real small-model errors (review point 2b):
#   off_by_one, wrong_operator, missing_return, wrong_variable,
#   wrong_edge (bad edge-case handling), swapped_branch, missing_case.
# --------------------------------------------------------------------------
CRITIQUE_TEMPLATES = {
    "off_by_one": [
        "First attempt:\n```python\n{buggy}\n```\nThat draft has an off-by-one error: {note}. The check `{failing}` fails because of it, so the boundary handling has to be corrected.",
        "First attempt:\n```python\n{buggy}\n```\nReviewing this, the loop/slice bound is off by one — {note}. Indeed, running the tests shows `{failing}` does not hold. Fix the bound and the behavior follows.",
        "First attempt:\n```python\n{buggy}\n```\nBug found: an off-by-one. {note}. Evidence: `{failing}` fails. The corrected version must include the endpoint properly.",
    ],
    "wrong_operator": [
        "First attempt:\n```python\n{buggy}\n```\nThe operator used here is wrong: {note}. That is why `{failing}` fails. Swapping in the correct operator fixes the logic.",
        "First attempt:\n```python\n{buggy}\n```\nProblem: the wrong comparison/arithmetic operator — {note}. The test `{failing}` exposes it. The fix is to use the right operator.",
        "First attempt:\n```python\n{buggy}\n```\nThis draft miscomputes because of an operator mistake: {note}. `{failing}` fails as a direct consequence. Correct the operator.",
    ],
    "wrong_edge": [
        "First attempt:\n```python\n{buggy}\n```\nEdge case handled incorrectly: {note}. The failing check `{failing}` is exactly that edge. Handle it explicitly.",
        "First attempt:\n```python\n{buggy}\n```\nThe main path is plausible but the edge case is wrong — {note}. `{failing}` fails. A correct solution must special-case or generalize this.",
        "First attempt:\n```python\n{buggy}\n```\nBoundary bug: {note}. Running the tests, `{failing}` breaks. Fix the edge handling.",
    ],
    "swapped_branch": [
        "First attempt:\n```python\n{buggy}\n```\nThe branches are effectively swapped: {note}. That makes `{failing}` fail. Put each result under the right condition.",
        "First attempt:\n```python\n{buggy}\n```\nLogic inversion: {note}. Evidence: `{failing}` does not hold. The two outcomes need to trade places.",
        "First attempt:\n```python\n{buggy}\n```\nThe conditional returns the wrong thing on each side — {note}. `{failing}` catches it. Swap the branch results.",
    ],
    "missing_case": [
        "First attempt:\n```python\n{buggy}\n```\nA required case is missing: {note}. The check `{failing}` fails because the input falls through. Add the missing case.",
        "First attempt:\n```python\n{buggy}\n```\nIncomplete: {note}. `{failing}` demonstrates the gap. The fix is to cover that case explicitly.",
        "First attempt:\n```python\n{buggy}\n```\nThis draft forgets a case — {note}. As a result `{failing}` fails. Cover it and the tests pass.",
    ],
    "missing_return": [
        "First attempt:\n```python\n{buggy}\n```\nThe computation is right but the result is never returned: {note}. Every check like `{failing}` fails because the function yields None. Add the return statement.",
        "First attempt:\n```python\n{buggy}\n```\nMissing return: {note}. The function computes the value and then drops it, so `{failing}` sees None. Return the computed value.",
        "First attempt:\n```python\n{buggy}\n```\nBug: no return statement — {note}. `{failing}` fails with None on the left side. The fix is one line: return the result.",
    ],
    "wrong_variable": [
        "First attempt:\n```python\n{buggy}\n```\nThe wrong variable is used: {note}. That is why `{failing}` fails — the computation reads or returns the wrong value. Use the correct variable.",
        "First attempt:\n```python\n{buggy}\n```\nVariable mix-up: {note}. The test `{failing}` exposes it. Swap in the intended variable.",
        "First attempt:\n```python\n{buggy}\n```\nIt operates on the wrong name — {note}. `{failing}` fails as a consequence. Correct the variable reference.",
    ],
}

# --------------------------------------------------------------------------
# Clean critiques (review point 2a): ~20% of Stage B examples have an
# ALREADY-CORRECT draft. The critique must explicitly say no bug was found;
# revise is a no-op. Slots: {ref}, {check}
# --------------------------------------------------------------------------
CRITIQUE_CLEAN = [
    "First attempt:\n```python\n{ref}\n```\nI reviewed this draft line by line against the requirements ({check}): the logic is correct, the edge cases are handled, and every test passes. No bug found — the draft stands as written.",
    "First attempt:\n```python\n{ref}\n```\nCareful review ({check}) turns up no defect: control flow, bounds, and return values are all right, and the tests pass. There is nothing to fix here.",
    "First attempt:\n```python\n{ref}\n```\nChecking this draft ({check}) — each step matches the specification and the full test suite passes. Verdict: no bug. Leaving the code unchanged.",
    "First attempt:\n```python\n{ref}\n```\nI looked for the usual suspects — off-by-one, wrong operator, missing return ({check}) — and found none. The draft is correct and all tests pass, so no revision is needed.",
]


def sha256_hex(s: str) -> str:
    return hashlib.sha256(s.encode("utf-8")).hexdigest()


def norm_instruction(s: str) -> str:
    return " ".join(s.lower().split())


def trivial_wrong_solutions(g: Gen) -> list:
    """Canonical wrong solutions used to guard against vacuous tests."""
    fn = g.func
    out = []
    if fn:
        out.append(f"def {fn}(*args, **kwargs):\n    return None")
        out.append(f"def {fn}(*args, **kwargs):\n    raise NotImplementedError")
    return out

	"""AGA-Code Stage B/C dataset framework.

	Everything here is original, procedurally generated, and execution-verified.
	No distillation, no third-party traces — correctness is established by
	RUNNING the code, never by copying anyone's answer.

	Execution harness semantics are byte-identical to aga_code.run_tests:
	harness = code + "\\n" + tests + "\\nprint('AGA_EXEC_OK')"
	pass = returncode == 0 and sentinel in stdout
	"""
	from __future__ import annotations

	import hashlib
	import random
	import subprocess
	import sys
	from dataclasses import dataclass, field

	MASTER_SEED = 1147
	SENTINEL = "AGA_EXEC_OK"


	# --------------------------------------------------------------------------
	# Execution harness (identical semantics to aga_code.run_tests)
	# --------------------------------------------------------------------------
	def run_code(code: str, tests: str, timeout: int = 5):
	"""Run code+tests in a subprocess. Returns (passed: bool, output: str)."""
	harness = code + "\n" + tests + "\nprint('AGA_EXEC_OK')"
	try:
	r = subprocess.run([sys.executable, "-c", harness],
	capture_output=True, timeout=timeout, text=True)
	ok = r.returncode == 0 and SENTINEL in (r.stdout or "")
	return ok, (r.stdout or "") + (r.stderr or "")
	except subprocess.TimeoutExpired:
	return False, "TIMEOUT"
	except Exception as e: # loud, never silent
	return False, f"HARNESS ERROR: {e}"


	def run_tests(code: str, tests: str, timeout: int = 5) -> bool:
	return run_code(code, tests, timeout)[0]


	class _ExecTimeout(BaseException):
	pass


	def _alarm_handler(signum, frame):
	raise _ExecTimeout()


	def exec_run(code: str, tests: str, timeout: float = 2.0) -> bool:
	"""Fast in-process execution gate with identical pass/fail semantics to
	run_tests (code + tests + sentinel print; pass = completes cleanly).

	Used for the high-volume build gates; the shipped data is additionally
	re-verified with the true subprocess harness (driver spot-check + the
	independent audit). BaseException timeout so mutant code cannot swallow
	it with a bare `except Exception`.
	"""
	import signal
	import contextlib
	import io
	src = code + "\n" + tests + "\nprint('AGA_EXEC_OK')"
	try:
	compiled = compile(src, "<gate>", "exec")
	except SyntaxError:
	return False
	old = signal.signal(signal.SIGALRM, _alarm_handler)
	signal.setitimer(signal.ITIMER_REAL, timeout)
	try:
	with contextlib.redirect_stdout(io.StringIO()):
	exec(compiled, {})
	return True
	except BaseException:
	return False
	finally:
	signal.setitimer(signal.ITIMER_REAL, 0)
	signal.signal(signal.SIGALRM, old)


	def first_failing_assert(code: str, tests: str, timeout: int = 5):
	"""Return the first single test line that fails against `code`, or None.

	Non-assert lines (e.g. fixture setup like `s = Stack()`) are prepended as
	setup context so the failing assertion reflects the real bug.
	"""
	setup = [ln for ln in tests.splitlines()
	if ln.strip() and not ln.strip().startswith("assert")]
	asserts = [ln for ln in tests.splitlines() if ln.strip().startswith("assert")]
	ctx = "\n".join(setup)
	for ln in asserts:
	ok, _ = run_code(code, ctx + "\n" + ln if ctx else ln, timeout)
	if not ok:
	return ln.strip()
	return None


	# --------------------------------------------------------------------------
	# Data structures
	# --------------------------------------------------------------------------
	@dataclass
	class Bug:
	code: str # buggy variant of the reference
	kind: str # off_by_one \| wrong_operator \| wrong_edge \| swapped_branch \| missing_case
	note: str # family-provided diagnosis detail (what is actually wrong)


	@dataclass
	class Gen:
	params: dict
	instruction: str # fully rendered instruction text
	phrasing: str # which instruction template produced it
	think_slots: dict # goal/approach/steps/edges/complexity
	reference: str # correct solution code
	tests: str # newline-joined assert statements
	difficulty: str # easy \| medium \| hard (per-instance)
	bugs: list = field(default_factory=list) # list[Bug]
	func: str = "" # primary function/class name


	@dataclass
	class Family:
	name: str
	difficulty: str # base: easy \| medium \| hard
	instruction_templates: list # list[str] with {slots}
	gen: object # callable(rng, hard: bool) -> Gen


	REGISTRY: list = []


	def family(name, difficulty, instruction_templates):
	def deco(fn):
	REGISTRY.append(Family(name=name, difficulty=difficulty,
	instruction_templates=instruction_templates, gen=fn))
	return fn
	return deco


	# --------------------------------------------------------------------------
	# Think skeletons — 12 varied phrasings filled from family-provided slots.
	# Slots: {goal} {approach} {steps} {edges} {complexity}
	# --------------------------------------------------------------------------
	THINK_SKELETONS = [
	"The task is to {goal}. My plan: {approach}. Concretely: {steps}. I have to be careful with {edges}. Overall this runs in {complexity}.",
	"Breaking it down: {goal}. The clean way to do this is {approach}. Steps: {steps}. Watch out for {edges}. Time complexity: {complexity}.",
	"Plan first. Goal: {goal}. Approach: {approach}. Then {steps}. Edge cases that could bite: {edges}. Expected complexity {complexity}.",
	"I need to {goal}. Strategy: {approach}. Walkthrough: {steps}. Before finalizing, check {edges}. This should be {complexity}.",
	"Let me think about {goal}. A direct approach is {approach}: {steps}. The risky parts are {edges}. Complexity is {complexity}.",
	"Decomposition: {goal}. Main idea: {approach}. Sub-steps: {steps}. Handle {edges} explicitly. Cost: {complexity}.",
	"For {goal}, I'll {approach}. The algorithm: {steps}. Non-obvious cases: {edges}. That gives {complexity}.",
	"Thinking it through: to {goal}, the natural method is {approach}. Sequence: {steps}. Don't forget {edges}. {complexity} overall.",
	"Approach for {goal}: {approach}. In order: {steps}. Special attention to {edges}. Complexity: {complexity}.",
	"First principles: {goal}. I'll {approach} because it is straightforward and correct: {steps}. Verify against {edges}. Runs in {complexity}.",
	"Target: {goal}. Method: {approach}. Implementation sketch: {steps}. Failure modes to avoid: {edges}. {complexity}.",
	"Here's my reasoning for {goal}: {approach} is the right tool. Steps: {steps}. Edge review: {edges}. Complexity {complexity}.",
	]

	# --------------------------------------------------------------------------
	# Critique templates per bug kind — filled with the family's diagnosis note
	# and the ACTUAL failing assertion observed by execution.
	# Bug kinds mirror real small-model errors (review point 2b):
	# off_by_one, wrong_operator, missing_return, wrong_variable,
	# wrong_edge (bad edge-case handling), swapped_branch, missing_case.
	# --------------------------------------------------------------------------
	CRITIQUE_TEMPLATES = {
	"off_by_one": [
	"First attempt:\n```python\n{buggy}\n```\nThat draft has an off-by-one error: {note}. The check `{failing}` fails because of it, so the boundary handling has to be corrected.",
	"First attempt:\n```python\n{buggy}\n```\nReviewing this, the loop/slice bound is off by one — {note}. Indeed, running the tests shows `{failing}` does not hold. Fix the bound and the behavior follows.",
	"First attempt:\n```python\n{buggy}\n```\nBug found: an off-by-one. {note}. Evidence: `{failing}` fails. The corrected version must include the endpoint properly.",
	],
	"wrong_operator": [
	"First attempt:\n```python\n{buggy}\n```\nThe operator used here is wrong: {note}. That is why `{failing}` fails. Swapping in the correct operator fixes the logic.",
	"First attempt:\n```python\n{buggy}\n```\nProblem: the wrong comparison/arithmetic operator — {note}. The test `{failing}` exposes it. The fix is to use the right operator.",
	"First attempt:\n```python\n{buggy}\n```\nThis draft miscomputes because of an operator mistake: {note}. `{failing}` fails as a direct consequence. Correct the operator.",
	],
	"wrong_edge": [
	"First attempt:\n```python\n{buggy}\n```\nEdge case handled incorrectly: {note}. The failing check `{failing}` is exactly that edge. Handle it explicitly.",
	"First attempt:\n```python\n{buggy}\n```\nThe main path is plausible but the edge case is wrong — {note}. `{failing}` fails. A correct solution must special-case or generalize this.",
	"First attempt:\n```python\n{buggy}\n```\nBoundary bug: {note}. Running the tests, `{failing}` breaks. Fix the edge handling.",
	],
	"swapped_branch": [
	"First attempt:\n```python\n{buggy}\n```\nThe branches are effectively swapped: {note}. That makes `{failing}` fail. Put each result under the right condition.",
	"First attempt:\n```python\n{buggy}\n```\nLogic inversion: {note}. Evidence: `{failing}` does not hold. The two outcomes need to trade places.",
	"First attempt:\n```python\n{buggy}\n```\nThe conditional returns the wrong thing on each side — {note}. `{failing}` catches it. Swap the branch results.",
	],
	"missing_case": [
	"First attempt:\n```python\n{buggy}\n```\nA required case is missing: {note}. The check `{failing}` fails because the input falls through. Add the missing case.",
	"First attempt:\n```python\n{buggy}\n```\nIncomplete: {note}. `{failing}` demonstrates the gap. The fix is to cover that case explicitly.",
	"First attempt:\n```python\n{buggy}\n```\nThis draft forgets a case — {note}. As a result `{failing}` fails. Cover it and the tests pass.",
	],
	"missing_return": [
	"First attempt:\n```python\n{buggy}\n```\nThe computation is right but the result is never returned: {note}. Every check like `{failing}` fails because the function yields None. Add the return statement.",
	"First attempt:\n```python\n{buggy}\n```\nMissing return: {note}. The function computes the value and then drops it, so `{failing}` sees None. Return the computed value.",
	"First attempt:\n```python\n{buggy}\n```\nBug: no return statement — {note}. `{failing}` fails with None on the left side. The fix is one line: return the result.",
	],
	"wrong_variable": [
	"First attempt:\n```python\n{buggy}\n```\nThe wrong variable is used: {note}. That is why `{failing}` fails — the computation reads or returns the wrong value. Use the correct variable.",
	"First attempt:\n```python\n{buggy}\n```\nVariable mix-up: {note}. The test `{failing}` exposes it. Swap in the intended variable.",
	"First attempt:\n```python\n{buggy}\n```\nIt operates on the wrong name — {note}. `{failing}` fails as a consequence. Correct the variable reference.",
	],
	}

	# --------------------------------------------------------------------------
	# Clean critiques (review point 2a): ~20% of Stage B examples have an
	# ALREADY-CORRECT draft. The critique must explicitly say no bug was found;
	# revise is a no-op. Slots: {ref}, {check}
	# --------------------------------------------------------------------------
	CRITIQUE_CLEAN = [
	"First attempt:\n```python\n{ref}\n```\nI reviewed this draft line by line against the requirements ({check}): the logic is correct, the edge cases are handled, and every test passes. No bug found — the draft stands as written.",
	"First attempt:\n```python\n{ref}\n```\nCareful review ({check}) turns up no defect: control flow, bounds, and return values are all right, and the tests pass. There is nothing to fix here.",
	"First attempt:\n```python\n{ref}\n```\nChecking this draft ({check}) — each step matches the specification and the full test suite passes. Verdict: no bug. Leaving the code unchanged.",
	"First attempt:\n```python\n{ref}\n```\nI looked for the usual suspects — off-by-one, wrong operator, missing return ({check}) — and found none. The draft is correct and all tests pass, so no revision is needed.",
	]


	def sha256_hex(s: str) -> str:
	return hashlib.sha256(s.encode("utf-8")).hexdigest()


	def norm_instruction(s: str) -> str:
	return " ".join(s.lower().split())


	def trivial_wrong_solutions(g: Gen) -> list:
	"""Canonical wrong solutions used to guard against vacuous tests."""
	fn = g.func
	out = []
	if fn:
	out.append(f"def {fn}(args, *kwargs):\n return None")
	out.append(f"def {fn}(args, *kwargs):\n raise NotImplementedError")
	return out