Datasets:
Download src/explicit_learning/interventions/substitute.py from sungguk/visual-answerability: direct link, hf CLI and curl.
- Browser
- Download file 11.3 kB
-
https://huggingface.co/datasets/sungguk/visual-answerability/resolve/main/src/explicit_learning/interventions/substitute.py
- Command line
-
hf download hf://datasets/sungguk/visual-answerability/src/explicit_learning/interventions/substitute.py
-
curl -L -o substitute.py https://huggingface.co/datasets/sungguk/visual-answerability/resolve/main/src/explicit_learning/interventions/substitute.py
11.3 kB
| """``SUBSTITUTE_V1`` — sample a new typed value for a node (§10.4). | |
| The substitute is drawn from the node's typed domain, deterministically seeded: | |
| * **numeric** (a plot point's ``y``, a table cell, a geometry measure) — a | |
| value within the source range and at the source precision, distinct from the | |
| current value. | |
| * **category** (a series / entity label) — a label that does not collide with | |
| any existing label. | |
| * **geometry** — after the substitute, the constraint graph must stay feasible | |
| (no contradiction); :mod:`feasibility` guards this. | |
| * **multiple choice** — when the problem is multiple-choice, the *new* answer | |
| the executor computes on the substituted world must match **exactly one** of | |
| the unchanged choices. A substitute that drives the answer out of the choice | |
| set (zero or >1 matches) is rejected (P5 gate: "objective choice outside | |
| unchanged choices reject"). | |
| The operator never labels the resulting state — it only validates that the | |
| substitute is *admissible*; P7 derives ``A_SAME`` / ``A_CHANGED`` by execution. | |
| """ | |
| from __future__ import annotations | |
| import hashlib | |
| import math | |
| from collections.abc import Iterator, Sequence | |
| from fractions import Fraction | |
| from typing import Any | |
| from ..dsl.ast import Dsl, Program, canonical_number_str, parse_number | |
| from ..executors.base import Executor, match_choice | |
| from ..ingest.base import Choice | |
| from ..sources.base import World | |
| from ..sources.world_ops import ( | |
| clone_world, | |
| find_point, | |
| find_row, | |
| find_series, | |
| find_stating_constraint, | |
| is_geometry, | |
| looks_numeric, | |
| node_kind, | |
| world_sha256, | |
| ) | |
| from .base import ( | |
| InterventionError, | |
| OperatorName, | |
| TransformedWorld, | |
| _set_value, | |
| transformed_hash, | |
| ) | |
| from .feasibility import FeasibilityError, is_feasible | |
| _NAME: OperatorName = "SUBSTITUTE_V1" | |
| class SubstituteOperator: | |
| """Substitute a seeded, typed, admissible new value onto ``target_node``.""" | |
| name = _NAME | |
| def apply( | |
| self, | |
| world: World, | |
| target_node: str, | |
| *, | |
| seed: int, | |
| dsl: Dsl, | |
| executor: Executor | None = None, | |
| program: Program | None = None, | |
| choices: Sequence[Choice] = (), | |
| answer_type: str = "", | |
| **_kwargs: Any, | |
| ) -> TransformedWorld: | |
| kind = node_kind(world, target_node) | |
| if kind is None: | |
| raise InterventionError(f"SUBSTITUTE_V1: target {target_node!r} is not a world node") | |
| value, desc = _sample_value(world, target_node, kind, seed) | |
| new = clone_world(world) | |
| if not _set_value(new, target_node, value): | |
| raise InterventionError( | |
| f"SUBSTITUTE_V1: target {target_node!r} has no mutable value site" | |
| ) | |
| if is_geometry(new): | |
| try: | |
| is_feasible(new, dsl=dsl) | |
| except FeasibilityError as exc: | |
| raise InterventionError( | |
| f"SUBSTITUTE_V1: infeasible substitute on {target_node!r}: {exc}" | |
| ) from exc | |
| # Multiple-choice admissibility: the new answer must be exactly one of | |
| # the unchanged choices (§10.4). This is a feasibility check, not a | |
| # state label — the operator still does not assign state. | |
| if answer_type == "multiple_choice": | |
| _assert_choice_admissible(new, executor, program, choices, target_node) | |
| return TransformedWorld( | |
| world=new, | |
| operator=_NAME, | |
| target_node=target_node, | |
| seed=seed, | |
| transformed_world_sha256=transformed_hash( | |
| operator=_NAME, | |
| target_node=target_node, | |
| seed=seed, | |
| source_world_sha256=world_sha256(world), | |
| transformed_world_sha256=world_sha256(new), | |
| hidden_nodes=frozenset(), | |
| substitute_value=value, | |
| ), | |
| hidden_nodes=frozenset(), | |
| substitute_value=value, | |
| description=desc, | |
| ) | |
| # --- admissibility --------------------------------------------------------- | |
| def _assert_choice_admissible( | |
| world: World, | |
| executor: Executor | None, | |
| program: Program | None, | |
| choices: Sequence[Choice], | |
| target_node: str, | |
| ) -> None: | |
| if executor is None or program is None: | |
| raise InterventionError( | |
| "SUBSTITUTE_V1: multiple-choice admissibility requires an executor and program" | |
| ) | |
| result = executor.execute(program, world=world) | |
| if result.status != "UNIQUE": | |
| raise InterventionError( | |
| f"SUBSTITUTE_V1: {target_node!r} substitute made the answer " | |
| f"non-unique ({result.status}); rejected" | |
| ) | |
| keys = match_choice(result.answer_value, choices, "multiple_choice") | |
| if len(keys) != 1: | |
| raise InterventionError( | |
| f"SUBSTITUTE_V1: new answer {result.answer_canonical!r} matches " | |
| f"{len(keys)} unchanged choices (must be exactly one); rejected" | |
| ) | |
| # --- typed sampling -------------------------------------------------------- | |
| def _sample_value(world: World, node_id: str, kind: str, seed: int) -> tuple[str, str]: | |
| if kind == "point": | |
| return _sample_point_y(world, node_id, seed) | |
| if kind == "series": | |
| return _sample_series_label(world, node_id, seed) | |
| if kind == "row": | |
| return _sample_row_cell(world, node_id, seed) | |
| if kind == "constraint": | |
| return _sample_geometry_measure(world, node_id, seed) | |
| if kind == "entity": | |
| return _sample_entity_label(world, node_id, seed) | |
| raise InterventionError(f"SUBSTITUTE_V1: no sampler for kind {kind!r}") | |
| def _sample_point_y(world: World, pid: str, seed: int) -> tuple[str, str]: | |
| point = find_point(world, pid) | |
| if point is None: | |
| raise InterventionError(f"point {pid!r} not found") | |
| sidx, _pidx, p = point | |
| siblings = world["series"][sidx].get("points", []) or [] | |
| values = [_frac(q.get("y", "0")) for q in siblings] | |
| cur = _frac(p.get("y", "0")) | |
| lo = min(values) if values else cur | |
| hi = max(values) if values else cur | |
| new = _seeded_in_range(seed, pid, "point_y", lo, hi, {cur}) | |
| return canonical_number_str(new), f"substitute point {pid} y → {new}" | |
| def _sample_row_cell(world: World, rid: str, seed: int) -> tuple[str, str]: | |
| row = find_row(world, rid) | |
| if row is None: | |
| raise InterventionError(f"row {rid!r} not found") | |
| _ridx, r = row | |
| cells = r.get("cells", {}) or {} | |
| if not cells: | |
| raise InterventionError(f"row {rid!r} has no cells") | |
| key = next(iter(cells)) | |
| cur = _frac(cells[key]) | |
| # Range across the same column in every row, if identifiable by shared key. | |
| col_vals = [ | |
| _frac(row_.get("cells", {}).get(key, cells[key])) | |
| for row_ in (world.get("rows", []) or []) | |
| if key in (row_.get("cells", {}) or {}) | |
| ] | |
| lo = min(col_vals) if col_vals else cur | |
| hi = max(col_vals) if col_vals else cur | |
| new = _seeded_in_range(seed, rid, f"cell:{key}", lo, hi, {cur}) | |
| return canonical_number_str(new), f"substitute row {rid} cell {key} → {new}" | |
| def _sample_series_label(world: World, sid: str, seed: int) -> tuple[str, str]: | |
| series = find_series(world, sid) | |
| if series is None: | |
| raise InterventionError(f"series {sid!r} not found") | |
| _sidx, s = series | |
| existing = {str(t.get("label", "")) for t in world.get("series", []) or []} | |
| cur = str(s.get("label", sid)) | |
| prefix = _label_prefix(cur) | |
| for n in _seeded_offsets(seed, sid, "label", span=1000): | |
| cand = f"{prefix}_{n}" | |
| if cand not in existing: | |
| return cand, f"substitute series {sid} label → {cand}" | |
| raise InterventionError(f"series {sid!r}: no non-colliding label found") | |
| def _sample_geometry_measure(world: World, entity: str, seed: int) -> tuple[str, str]: | |
| found = find_stating_constraint(world, entity) | |
| if found is None: | |
| raise InterventionError(f"no stating constraint for {entity!r}") | |
| _cidx, c = found | |
| args = list(c.get("args") or []) | |
| pred = c.get("predicate") | |
| if pred == "MeasureOf" and len(args) >= 2: | |
| cur = _frac(args[1]) | |
| elif pred == "Equals" and len(args) == 2: | |
| cur = _frac(args[1]) if looks_numeric(args[1]) else _frac(args[0]) | |
| else: | |
| raise InterventionError(f"stating constraint for {entity!r} is not numeric") | |
| # Geometry measures are unbounded; sample a distinct integer near the source | |
| # (feasibility, not range, is the binding constraint here). | |
| new = cur + Fraction(next(_seeded_offsets(seed, entity, "measure", span=200)) - 100) | |
| return canonical_number_str(new), f"substitute measure {entity} → {new}" | |
| def _sample_entity_label(world: World, eid: str, seed: int) -> tuple[str, str]: | |
| entity = next((e for e in world.get("entities", []) or [] if e.get("id") == eid), None) | |
| if entity is None: | |
| raise InterventionError(f"entity {eid!r} not found") | |
| cur = str(entity.get("label", entity.get("id", eid))) | |
| existing = {str(e.get("label", e.get("id", ""))) for e in world.get("entities", []) or []} | |
| prefix = _label_prefix(cur) | |
| for n in _seeded_offsets(seed, eid, "elabel", span=1000): | |
| cand = f"{prefix}_{n}" | |
| if cand not in existing: | |
| return cand, f"substitute entity {eid} label → {cand}" | |
| raise InterventionError(f"entity {eid!r}: no non-colliding label found") | |
| # --- seeded helpers (deterministic, Math.random-free) ---------------------- | |
| def _seeded_in_range( | |
| seed: int, tag: str, kind: str, lo: Fraction, hi: Fraction, exclude: set[Fraction] | |
| ) -> Fraction: | |
| """A deterministic integer in ``[lo, hi]`` (extended by ±1 when degenerate), | |
| not in ``exclude``, at source integer precision.""" | |
| lo_i = int(lo) | |
| hi_i = int(hi) | |
| if hi_i < lo_i: | |
| lo_i, hi_i = hi_i, lo_i | |
| if hi_i - lo_i < 1: # degenerate range: relax by ±1 to allow a distinct value | |
| lo_i -= 1 | |
| hi_i += 1 | |
| span = hi_i - lo_i + 1 | |
| # An affine full-cycle permutation needs at most ``len(exclude) + 1`` | |
| # probes to find a non-excluded integer. Never enumerate a numeric range: | |
| # PlotQA values can make ``span`` billions wide. | |
| for probe, n in enumerate(_seeded_offsets(seed, tag, kind, span=span)): | |
| if probe > len(exclude): | |
| break | |
| cand = Fraction(lo_i + (n % (hi_i - lo_i + 1))) | |
| if cand not in exclude: | |
| return cand | |
| raise InterventionError(f"no in-range substitute for {tag!r} ({kind})") | |
| def _seeded_offsets(seed: int, tag: str, kind: str, *, span: int) -> Iterator[int]: | |
| """Yield a deterministic O(1)-memory permutation of ``[0, span)``.""" | |
| span = max(span, 1) | |
| digest = hashlib.sha256(f"{seed}|{tag}|{kind}|affine-v1".encode()).digest() | |
| start = int.from_bytes(digest[:16], "big") % span | |
| step = int.from_bytes(digest[16:], "big") % span | |
| if step == 0: | |
| step = 1 | |
| while math.gcd(step, span) != 1: | |
| step = (step + 1) % span | |
| if step == 0: | |
| step = 1 | |
| for index in range(span): | |
| yield (start + index * step) % span | |
| def _frac(value: Any) -> Fraction: | |
| return parse_number(value) | |
| def _label_prefix(label: str) -> str: | |
| return label.rstrip("0123456789_") or label or "node" | |
| __all__ = ["SubstituteOperator"] | |