84db129629
GSM8K labels containers/regions with a trailing single-letter or short-numeric
label ('Jar A has 28 marbles', 'Section G has 10 cars', 'District 2 has 19
voters'); the initial-possession entity slot captured only 'Jar' and the label
broke the match. Adds a separate sibling pattern _INITIAL_HAS_LABELED_RE
(mirroring ADR-0136.S.4 localisation) that REQUIRES the label, so the global
_ENTITY is unchanged and bare subjects yield no duplicate candidate.
- Composes with ADR-0193 aggregate question: 'Jar A has 28 marbles. Jar B has
12 marbles. How many marbles are there in total?' -> 40.0.
- 0 real-corpus metric flip (honest substrate): the one real multi-container
aggregate additionally needs comparative + multiplicative + lowercase-ref.
- wrong=0 HOLDS full corpus (7,473 q); train_sample byte-identical 4/46/0;
synthetic-registry capability-axis gate + G5 lane green; smoke 67 passed.
- Label bounded by the possession verb: multi-word nouns ('Jar Apple') do NOT
match. wrong=0 held downstream by completeness + round-trip + disagreement.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
81 lines
3.6 KiB
Python
81 lines
3.6 KiB
Python
"""ADR-0194 — labeled-container subject entity shape.
|
|
|
|
GSM8K routinely labels containers/regions with a trailing single-letter or
|
|
short-numeric label: "Jar A has 28 marbles.", "Section G has 15 rows.",
|
|
"District 2 has 19 voters.". The initial-possession parser's entity slot
|
|
(``_ENTITY = (?:[A-Z]\\w+|[Tt]he\\s+\\w+)``) captures only "Jar" and then
|
|
expects the possession verb, so the label breaks the match and the statement
|
|
produces no candidate.
|
|
|
|
This adds a SEPARATE sibling pattern ``_INITIAL_HAS_LABELED_RE`` (mirroring
|
|
ADR-0136.S.4's ``_INITIAL_HAS_INDEF_RE`` localisation) that REQUIRES a label,
|
|
so it never duplicates the bare-subject main pattern. The global ``_ENTITY``
|
|
is unchanged. wrong=0 is held downstream by the completeness guard
|
|
(ADR-0191) + round-trip + branch disagreement — the label widening only makes
|
|
a statement *parse*; a mis-parse leaves quantities uncovered and refuses.
|
|
|
|
SUBSTRATE: 0 real-corpus metric flip (the one real multi-container aggregate,
|
|
"Jar A has 28 marbles. Jar B has 12 more than jar A. Jar C has twice as many
|
|
as jar B. ...altogether?", additionally needs comparative + multiplicative
|
|
reading). Its value is the entity-shape generalisation + that it composes
|
|
with the ADR-0193 aggregate question (proven below).
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
import pytest
|
|
|
|
from generate.math_candidate_parser import extract_initial_candidates
|
|
from generate.math_candidate_graph import parse_and_solve
|
|
|
|
|
|
# --- Labeled-container subjects now parse as initial possessions -----------
|
|
@pytest.mark.parametrize("sentence,entity,value,unit", [
|
|
("Jar A has 28 marbles.", "Jar A", 28.0, "marbles"),
|
|
("Box B has 15 marbles.", "Box B", 15.0, "marbles"),
|
|
("Section G has 10 cars.", "Section G", 10.0, "cars"),
|
|
("District 2 has 19 voters.", "District 2", 19.0, "voters"),
|
|
("Tank 1 has 40 liters.", "Tank 1", 40.0, "liters"),
|
|
])
|
|
def test_labeled_container_parses(sentence, entity, value, unit) -> None:
|
|
cands = extract_initial_candidates(sentence)
|
|
assert len(cands) == 1, f"expected one candidate for {sentence!r}"
|
|
c = cands[0]
|
|
assert c.initial.entity == entity
|
|
assert c.initial.quantity.value == value
|
|
assert c.initial.quantity.unit == unit
|
|
|
|
|
|
# --- No duplicate candidates for bare (unlabeled) subjects -----------------
|
|
def test_bare_subject_single_candidate_unchanged() -> None:
|
|
"""The labeled pattern requires a label, so 'Jamie has 28 marbles'
|
|
still yields exactly one candidate (from the main pattern only)."""
|
|
cands = extract_initial_candidates("Jamie has 28 marbles.")
|
|
assert len(cands) == 1
|
|
assert cands[0].initial.entity == "Jamie"
|
|
|
|
|
|
# --- The label must not swallow a following content word -------------------
|
|
@pytest.mark.parametrize("sentence", [
|
|
"Jar Apple has 5 marbles.", # 'Apple' is not a single-letter label
|
|
"Box Set has 12 items.", # 'Set' is not a label
|
|
])
|
|
def test_multiword_noun_not_a_label(sentence: str) -> None:
|
|
cands = extract_initial_candidates(sentence)
|
|
assert cands == [], f"{sentence!r} must not parse as a labeled container"
|
|
|
|
|
|
# --- Composes with the ADR-0193 aggregate question -------------------------
|
|
def test_composes_with_aggregate_question() -> None:
|
|
res = parse_and_solve(
|
|
"Jar A has 28 marbles. Jar B has 12 marbles. "
|
|
"How many marbles are there in total?"
|
|
)
|
|
assert res.answer == 40.0, res.refusal_reason
|
|
|
|
|
|
def test_composes_three_containers() -> None:
|
|
res = parse_and_solve(
|
|
"Jar A has 5 marbles. Jar B has 3 marbles. Jar C has 2 marbles. "
|
|
"How many marbles are there altogether?"
|
|
)
|
|
assert res.answer == 10.0, res.refusal_reason
|