File size: 13,075 Bytes
15d1f1d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 | #!/usr/bin/env python3
"""Final prompt-annotation causal-isolation repair layered on RC5J-final.
RC5K changes exactly one formal component: prompt-scoped annotation. The
offline full-GT annotation component, the other five components in both
scopes, the full-GT reward, geometry matching, anti-hack gates, and benchmark
remain RC5J-final.
Prompt native dimensions are partitioned with one joint lexicographic
assignment: maximise selected matches, then neutralise as many unselected GT
matches as possible, then minimise pair cost. Neutral matches are neither
precision credit nor spam. Remaining native dimensions and every
dimension-like TEXT/MTEXT entity are spam. This removes full-GT inventory
leakage while preserving the intended neutral treatment of correct dimensions
outside the prompt mask.
"""
from __future__ import annotations
import math
from pathlib import Path
from typing import Any, Mapping
import numpy as np
try:
from . import eval_geometric052_rc5j_final as RC5J
from . import prompt_annotation_precision_candidate_v2 as PRECISION
except ImportError: # direct evaluator import
import eval_geometric052_rc5j_final as RC5J
import prompt_annotation_precision_candidate_v2 as PRECISION
METRIC_VERSION = "v5.2-rc5k-final"
SCHEMA_VERSION = "edp-geometric-evaluator-v5.2-rc5k-final"
ENGINEERING_THRESHOLDS = dict(RC5J.ENGINEERING_THRESHOLDS)
PHYSICAL_SUPPORTED_SEMANTICS = RC5J.PHYSICAL_SUPPORTED_SEMANTICS
MAX_MATCHED_OFFSET_DISPERSION_FRACTION = (
RC5J.MAX_MATCHED_OFFSET_DISPERSION_FRACTION
)
RC3 = RC5J.RC3
BASE = RC5J.BASE
def _prompt_annotation_component(
context: Mapping[str, Any],
sidecar: Mapping[str, Any] | None,
) -> dict[str, Any]:
"""Score prompt targets with causally isolated native-DIM precision."""
contract, selected_contract = RC3._scope_contract(sidecar, "prompt")
mask = selected_contract.get("annotation_mask", [])
if not selected_contract.get("annotation_eligible"):
return BASE._component(
None,
False,
"prompt_annotation_mask_ineligible",
scope="prompt",
contract_errors=contract.get("errors", []),
reason_codes=selected_contract.get("reason_codes", []),
)
targets = RC3._dimension_targets(context, sidecar, contract)
selected_spec_indices = {
index for index, selected in enumerate(mask) if selected
}
selected_target_mask = [
target["spec_index"] in selected_spec_indices for target in targets
]
selected_targets = [
target
for target, selected in zip(targets, selected_target_mask)
if selected
]
if not selected_targets:
return BASE._component(
None,
False,
"no_valid_masked_target_dimensions",
scope="prompt",
)
pred_dims = list((context.get("raw_pred") or {}).get("dims", []))
details: dict[tuple[int, int], dict[str, Any]] = {}
pair_matrix: list[list[dict[str, Any]]] = []
for pred_index, pred_dim in enumerate(pred_dims):
row: list[dict[str, Any]] = []
for target_index, target in enumerate(targets):
detail = RC3._dimension_pair_detail(context, pred_dim, target)
details[(pred_index, target_index)] = detail
row.append(
{
"family_ok": detail.get("family_ok"),
"binding_distance_fraction": detail.get(
"binding_distance_fraction"
),
"predicted_value": detail.get("predicted_value"),
}
)
pair_matrix.append(row)
pred_meta = context.get("pred_meta", {})
pred_text = int(pred_meta.get("dimension_like_text", 0) or 0)
try:
precision = PRECISION.prompt_annotation_precision_candidate_v2(
pair_matrix,
selected_target_mask=selected_target_mask,
target_values=[abs(float(target["spec"]["value"])) for target in targets],
target_ids=[str(target["spec"]["id"]) for target in targets],
predicted_text_count=pred_text,
max_binding_fraction=RC3.MAX_DIMENSION_BINDING_FRACTION,
max_value_log_error=RC3.MAX_DIMENSION_VALUE_LOG_ERROR,
)
except (KeyError, TypeError, ValueError) as exc:
return BASE._component(
None,
False,
"prompt_annotation_precision_contract_invalid",
scope="prompt",
contract_error=f"{type(exc).__name__}: {exc}",
precision_policy=PRECISION.POLICY_VERSION,
)
if not precision.get("eligible"):
return BASE._component(
None,
False,
str(precision.get("reason") or "prompt_annotation_precision_ineligible"),
scope="prompt",
precision_policy=PRECISION.POLICY_VERSION,
precision_diagnostics=precision,
)
selected_pair_by_target = {
int(pair["target_index"]): int(pair["pred_index"])
for pair in precision["selected_pairs"]
}
target_position = {
target["spec_index"]: index for index, target in enumerate(targets)
}
item_rows: list[dict[str, Any]] = []
item_scores: list[float] = []
presence_values: list[float] = []
type_values: list[float] = []
value_values: list[float] = []
binding_values: list[float] = []
for target in selected_targets:
position = target_position[target["spec_index"]]
pred_index = selected_pair_by_target.get(position)
spec = target["spec"]
if pred_index is None:
item_scores.append(0.0)
presence_values.append(0.0)
type_values.append(0.0)
value_values.append(0.0)
binding_values.append(0.0)
item_rows.append(
{
"id": spec["id"],
"gt_dimension_index": target["gt_index"],
"pred_dimension_index": None,
"role": target.get("role"),
"presence": 0.0,
"type": 0.0,
"display_value": 0.0,
"binding": 0.0,
"score": 0.0,
"reason": "missing_or_semantically_incompatible_dimension",
}
)
continue
detail = details[(pred_index, position)]
value_log_error = float(detail["value_log_error"])
binding_distance = float(detail["binding_distance_fraction"])
value_quality = math.exp(-value_log_error / 0.10)
binding_quality = math.exp(-binding_distance / 0.03)
item_score = 0.35 + 0.20 + 0.30 * value_quality + 0.15 * binding_quality
item_scores.append(item_score)
presence_values.append(1.0)
type_values.append(1.0)
value_values.append(value_quality)
binding_values.append(binding_quality)
item_rows.append(
{
"id": spec["id"],
"gt_dimension_index": target["gt_index"],
"pred_dimension_index": pred_index,
"role": target.get("role"),
"presence": 1.0,
"type": 1.0,
"display_value": BASE._round(value_quality),
"binding": BASE._round(binding_quality),
"target_display_value": BASE._round(detail["target_value"]),
"predicted_display_value": BASE._round(detail["predicted_value"]),
"display_value_source": detail["value_source"],
"display_value_relative_error": BASE._round(
detail["value_relative_error"]
),
"binding_distance_fraction": BASE._round(binding_distance),
"score": BASE._round(item_score),
"reason": (
"ok" if item_score >= 0.90 else "annotation_field_mismatch"
),
}
)
target_macro = float(np.mean(item_scores))
target_worst = min(item_scores)
target_quality = 0.50 * target_macro + 0.50 * target_worst
predicted_precision = float(precision["precision"])
score = BASE._harmonic(target_quality, predicted_precision)
selected_count = int(precision["n_selected_matched_native"])
neutral_count = int(precision["n_neutralized_unselected_native"])
unmatched_count = int(precision["n_unmatched_predicted_native"])
return BASE._component(
score,
True,
(
"ok"
if score >= ENGINEERING_THRESHOLDS["annotation"]
else "annotation_recall_precision_or_field_mismatch"
),
scope="prompt",
n_selected_target=len(selected_targets),
n_total_gt_dimension_targets=len(targets),
n_predicted_native=len(pred_dims),
n_predicted_dimension_like_text=pred_text,
n_selected_matched_native=selected_count,
n_neutralized_unselected_native=neutral_count,
n_unmatched_predicted_native=unmatched_count,
n_matched_predicted_native=selected_count + neutral_count,
predicted_precision=BASE._round(predicted_precision),
effective_predicted_inventory=int(
precision["effective_predicted_inventory"]
),
empty_effective_inventory=bool(precision["empty_effective_inventory"]),
presence_recall=BASE._round(float(np.mean(presence_values))),
type_macro=BASE._round(float(np.mean(type_values))),
display_value_macro=BASE._round(float(np.mean(value_values))),
binding_macro=BASE._round(float(np.mean(binding_values))),
target_quality=BASE._round(target_quality),
item_macro=BASE._round(target_macro),
item_worst=BASE._round(target_worst),
items=item_rows,
matching_policy=PRECISION.POLICY_VERSION,
selected_targets_have_lexicographic_priority=True,
unselected_gt_dimensions_are_precision_neutral=True,
unselected_matches_are_neither_credit_nor_spam=True,
dimension_like_text_policy="always_unmatched_without_item_level_verifier",
reads_actual_geometry=False,
)
def score_dxf_pair(
pred: str | Path,
gt: str | Path,
sidecar: Mapping[str, Any] | None = None,
category: str | None = None,
**kwargs: Any,
) -> dict[str, Any]:
"""Run RC5J-final with only its prompt annotation callback replaced."""
# Reuse RC5J's RLock so concurrent RC5J/RC5K calls cannot observe a
# half-patched RC3 module. RC5J re-enters this lock during its own wrapper.
with RC5J._PATCH_LOCK:
original_annotation = RC3._annotation_component
def scoped_annotation(
context: Mapping[str, Any],
candidate_sidecar: Mapping[str, Any] | None,
*,
scope: str,
) -> dict[str, Any]:
if scope == "prompt":
return _prompt_annotation_component(context, candidate_sidecar)
return original_annotation(
context,
candidate_sidecar,
scope=scope,
)
RC3._annotation_component = scoped_annotation
try:
result = RC5J.score_dxf_pair(
pred,
gt,
sidecar=sidecar,
category=category,
**kwargs,
)
finally:
RC3._annotation_component = original_annotation
result["schema_version"] = SCHEMA_VERSION
result["metric_version"] = METRIC_VERSION
result.setdefault("metric_definition", {})[
"prompt_annotation_precision_policy"
] = {
"policy_version": PRECISION.POLICY_VERSION,
"changed_scope": "online_reward_prompt_masked.annotation_only",
"assignment_objective": [
"maximise_selected_matches",
"maximise_unselected_neutral_matches",
"minimise_pair_cost",
],
"precision": "S/(S+X+Q)",
"unselected_native_matches": "neutral",
"dimension_like_text": "unmatched_without_item_level_verifier",
"zero_unselected_value": "absolute_tolerance_neutralisation",
"zero_selected_value": "fail_closed_ineligible",
"full_gt_scope_changed": False,
"other_component_definitions_changed": False,
}
result.setdefault("diagnostics", {})["rc5k_prompt_annotation_precision"] = {
"applied": True,
"policy_version": PRECISION.POLICY_VERSION,
"changed_scope": "prompt_annotation_and_dependent_prompt_reward_only",
"full_gt_scope_changed": False,
"other_five_components_changed": False,
}
release = result.get("diagnostics", {}).get("release_status")
if isinstance(release, dict):
release["candidate"] = METRIC_VERSION
return result
__all__ = [
"ENGINEERING_THRESHOLDS",
"MAX_MATCHED_OFFSET_DISPERSION_FRACTION",
"METRIC_VERSION",
"PHYSICAL_SUPPORTED_SEMANTICS",
"SCHEMA_VERSION",
"_prompt_annotation_component",
"score_dxf_pair",
]
|