Datasets:
File size: 17,774 Bytes
e1ced61 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 | """Grouped data and pinned-TRL contracts for EVI-PO.
EVI-PO never exposes intervention labels to the model. The labels below are
loader-side relationships used to bind four independently rendered prompts:
``FULL(I0, a)``, ``CONTROL(Ic, a)``, and
``MISSING(Im, <UNANSWERABLE>)`` are mandatory. ``SUBSTITUTE(Id, a')`` is
optional because real-image sources such as GQA do not have a certified
answer-changing edit.
The evidence target is not a changed-pixel heuristic. It is the executor's
dependency-node trace projected through the renderer's lossless node map. The
trainer later projects that pixel ownership mask onto the processor's visual
token grid.
"""
from __future__ import annotations
import copy
import inspect
import json
from collections import defaultdict
from collections.abc import Iterable, Mapping
from dataclasses import dataclass
from pathlib import Path
from typing import Any
from ..certificates.build import certificate_id
from ..hashing import canonical_json_hash
from ..renderers.node_map import decode_uint32_png
from .answers import UNANSWERABLE_TOKEN
from .slots import ComparisonSlot
from .targets import ours_target, target_text
EVI_PO_ADAPTER_VERSION = "evi_po_on_trl_1.9.1_v1"
EVI_RELATION_ORDER = ("FULL", "CONTROL", "SUBSTITUTE", "MISSING")
EVI_REQUIRED_RELATIONS = ("FULL", "CONTROL", "MISSING")
_RELATION_CONTRACT = {
"FULL": ("FULL", "POSITIVE"),
"CONTROL": ("A_SAME", "CONTROL"),
"SUBSTITUTE": ("A_CHANGED", "SUBSTITUTE"),
"MISSING": ("U_MISSING", "TARGET_EVIDENCE"),
}
class EVIContractError(ValueError):
"""Raised when grouped data cannot support the registered EVI-PO loss."""
@dataclass(frozen=True)
class EVIGroupContract:
"""One complete, question-identical four-view EVI-PO relationship."""
group_id: str
base_id: str
question: str
answer_type: str
choices: tuple[dict[str, Any], ...]
full_answer: str
substitute_answer: str | None
relations: tuple[tuple[str, ComparisonSlot], ...]
@property
def candidates(self) -> tuple[str, ...]:
"""Finite candidate support used by the FULL/CONTROL JS term."""
values = [self.full_answer]
if self.substitute_answer is not None:
values.append(self.substitute_answer)
values.append(UNANSWERABLE_TOKEN)
return tuple(values)
def relation(self, name: str) -> ComparisonSlot:
try:
return dict(self.relations)[name]
except KeyError:
raise EVIContractError(f"group {self.group_id}: missing EVI relation {name}") from None
def build_evi_group_contracts(
slots: Iterable[ComparisonSlot],
) -> dict[str, EVIGroupContract]:
"""Group persisted comparison slots and enforce the EVI relationship.
``U_INVALID`` and any future independently scored views may remain in the
comparison schedule, but they do not replace one of the four directional
relationship members.
"""
by_group: dict[str, list[ComparisonSlot]] = defaultdict(list)
for slot in slots:
by_group[slot.group_id].append(slot)
if not by_group:
raise EVIContractError("cannot build EVI groups from an empty slot sequence")
grouped: dict[str, EVIGroupContract] = {}
for group_id, group_slots in by_group.items():
first = group_slots[0]
relation_slots: dict[str, ComparisonSlot] = {}
first_full_hash = canonical_json_hash(first.full_view)
first_choices_hash = canonical_json_hash(list(first.choices))
for slot in group_slots:
if (
slot.base_id != first.base_id
or slot.question != first.question
or slot.answer_type != first.answer_type
or canonical_json_hash(list(slot.choices)) != first_choices_hash
or canonical_json_hash(slot.full_view) != first_full_hash
or target_text(slot.full_answer) != target_text(first.full_answer)
):
raise EVIContractError(
f"group {group_id}: prompt, answer, or FULL-view identity drift"
)
state = str(slot.selected_view.get("state", ""))
role = str(slot.selected_view.get("role", ""))
for relation, expected in _RELATION_CONTRACT.items():
if (state, role) != expected:
continue
if relation in relation_slots:
raise EVIContractError(
f"group {group_id}: multiple {relation} relationship views"
)
relation_slots[relation] = slot
missing = [name for name in EVI_REQUIRED_RELATIONS if name not in relation_slots]
if missing:
raise EVIContractError(
f"group {group_id}: incomplete EVI relationship; missing {', '.join(missing)}"
)
full_answer = target_text(first.full_answer)
if ours_target(relation_slots["FULL"].selected_view, first.full_answer) != full_answer:
raise EVIContractError(f"group {group_id}: FULL target differs from group answer")
if ours_target(relation_slots["CONTROL"].selected_view, first.full_answer) != full_answer:
raise EVIContractError(f"group {group_id}: CONTROL target differs from FULL")
substitute_answer: str | None = None
if "SUBSTITUTE" in relation_slots:
substitute_answer = ours_target(
relation_slots["SUBSTITUTE"].selected_view,
first.full_answer,
)
if substitute_answer in {full_answer, UNANSWERABLE_TOKEN}:
raise EVIContractError(
f"group {group_id}: SUBSTITUTE target is not a distinct concrete answer"
)
if (
ours_target(relation_slots["MISSING"].selected_view, first.full_answer)
!= UNANSWERABLE_TOKEN
):
raise EVIContractError(f"group {group_id}: MISSING target is not unanswerable")
grouped[group_id] = EVIGroupContract(
group_id=group_id,
base_id=first.base_id,
question=first.question,
answer_type=first.answer_type,
choices=tuple(copy.deepcopy(first.choices)),
full_answer=full_answer,
substitute_answer=substitute_answer,
relations=tuple(
(name, relation_slots[name])
for name in EVI_RELATION_ORDER
if name in relation_slots
),
)
return grouped
def _load_object(path: Path, label: str) -> dict[str, Any]:
try:
value = json.loads(path.read_text(encoding="utf-8"))
except (OSError, json.JSONDecodeError) as exc:
raise EVIContractError(f"cannot read {label} {path}: {exc}") from exc
if not isinstance(value, dict):
raise EVIContractError(f"{label} must be an object: {path}")
return value
def _inside(root: Path, path: Path, label: str) -> Path:
resolved_root = root.resolve()
resolved = path.resolve()
try:
resolved.relative_to(resolved_root)
except ValueError as exc:
raise EVIContractError(f"{label} escapes dataset root: {path}") from exc
return resolved
def evidence_spec_for_view(
view: Mapping[str, Any],
*,
dataset_root: str | Path,
) -> dict[str, Any]:
"""Resolve and verify the certificate/node-map inputs for evidence loss.
Every returned spec carries ``evidence_available``: ``True`` when the
certified trace projects to at least one source pixel, ``False`` when the
artifacts are well-formed but the dependency node has no visible footprint
(a structural property of some GQA/CLEVR renders — the executor depends on
a node that occupies zero pixels in the lossless node map). A spec with
``evidence_available=False`` is a graceful per-group skip signal: the
trainer drops that group's evidence term while keeping its answer and
direction terms, so every arm still trains on the same slot set.
A *total* absence of evidence artifacts (no node map, no source mask, no
annotation boxes) still fails closed — that is a data defect, not a
renderable-but-empty trace.
"""
root = Path(dataset_root).resolve()
images = view.get("images")
if not isinstance(images, list) or len(images) != 1 or not isinstance(images[0], Mapping):
raise EVIContractError("EVI evidence projection requires exactly one image per view")
raw_image = images[0].get("path")
if not isinstance(raw_image, str) or not raw_image:
raise EVIContractError("EVI view image path is missing")
relative = Path(raw_image)
if relative.is_absolute() or ".." in relative.parts or "\\" in raw_image:
raise EVIContractError(f"unsafe EVI image path: {raw_image!r}")
image_path = _inside(root, root / relative, "EVI image")
if not image_path.is_file():
raise EVIContractError(f"EVI image is missing: {image_path}")
image_width = int(images[0].get("width", 0))
image_height = int(images[0].get("height", 0))
if image_width <= 0 or image_height <= 0:
raise EVIContractError("EVI image dimensions are invalid")
artifact_dir = image_path.parent
certificate_path = _inside(root, artifact_dir / "certificate.json", "certificate")
manifest_path = _inside(root, artifact_dir / "render_manifest.json", "render manifest")
node_map_path = _inside(root, artifact_dir / "node_map.uint32.png", "node map")
structured_paths = (certificate_path, manifest_path, node_map_path)
if all(path.is_file() for path in structured_paths):
certificate = _load_object(certificate_path, "certificate")
manifest = _load_object(manifest_path, "render manifest")
if (
certificate.get("certificate_id") != view.get("certificate_id")
or certificate_id(certificate) != certificate.get("certificate_id")
or certificate.get("view_id") != view.get("view_id")
):
raise EVIContractError(f"view {view.get('view_id')}: certificate identity mismatch")
program = certificate.get("program")
dependency_ids = (
program.get("dependency_node_ids") if isinstance(program, Mapping) else None
)
if (
not isinstance(dependency_ids, list)
or not dependency_ids
or any(not isinstance(node_id, str) or not node_id for node_id in dependency_ids)
):
raise EVIContractError(
f"view {view.get('view_id')}: certificate has no executor dependency trace"
)
node_table = manifest.get("node_table")
if (
not isinstance(node_table, list)
or not node_table
or node_table[0] != "__background__"
or any(not isinstance(node_id, str) for node_id in node_table)
):
raise EVIContractError(f"view {view.get('view_id')}: malformed node table")
try:
owners = decode_uint32_png(node_map_path.read_bytes())
import numpy as np
owner_indices = [
node_table.index(node_id) for node_id in dependency_ids if node_id in node_table
]
evidence_pixels = int(np.isin(owners, owner_indices).sum()) if owner_indices else 0
height, width = (int(value) for value in owners.shape)
except (OSError, TypeError, ValueError) as exc:
raise EVIContractError(
f"view {view.get('view_id')}: cannot decode executor evidence mask: {exc}"
) from exc
if (
(width, height) != (image_width, image_height)
or manifest.get("width") != width
or manifest.get("height") != height
):
raise EVIContractError(f"view {view.get('view_id')}: node-map/image dimensions differ")
return {
"schema_version": 1,
"source": "executor_dependency_nodes_projected_through_node_map",
"certificate_path": str(certificate_path),
"certificate_id": str(certificate["certificate_id"]),
"render_manifest_path": str(manifest_path),
"render_manifest_sha256": view.get("render_manifest_sha256"),
"node_map_path": str(node_map_path),
"node_map_sha256": manifest.get("node_map_sha256"),
"dependency_node_ids": list(dependency_ids),
"node_table": list(node_table),
"width": width,
"height": height,
"positive_pixel_count": evidence_pixels,
"evidence_available": evidence_pixels > 0,
}
if any(path.exists() for path in structured_paths):
raise EVIContractError(
f"view {view.get('view_id')}: structured evidence artifacts are incomplete"
)
# Real-image sources can carry a source-native binary evidence mask.
raw_mask_path = view.get("evidence_mask_path", images[0].get("evidence_mask_path"))
raw_mask_sha = view.get("evidence_mask_sha256", images[0].get("evidence_mask_sha256"))
if isinstance(raw_mask_path, str) and raw_mask_path:
relative_mask = Path(raw_mask_path)
if relative_mask.is_absolute() or ".." in relative_mask.parts or "\\" in raw_mask_path:
raise EVIContractError(f"unsafe evidence-mask path: {raw_mask_path!r}")
mask_path = _inside(root, root / relative_mask, "evidence mask")
if not mask_path.is_file():
raise EVIContractError(f"view {view.get('view_id')}: evidence mask is missing")
try:
import numpy as np
from PIL import Image
with Image.open(mask_path) as mask_image:
mask = np.asarray(mask_image.convert("L")) > 0
height, width = (int(value) for value in mask.shape)
evidence_pixels = int(mask.sum())
except (OSError, TypeError, ValueError) as exc:
raise EVIContractError(
f"view {view.get('view_id')}: cannot decode explicit evidence mask: {exc}"
) from exc
if (width, height) != (image_width, image_height):
raise EVIContractError(
f"view {view.get('view_id')}: evidence-mask/image dimensions differ"
)
return {
"schema_version": 1,
"source": "source_annotation_explicit_mask",
"mask_path": str(mask_path),
"mask_sha256": raw_mask_sha,
"width": width,
"height": height,
"positive_pixel_count": evidence_pixels,
"evidence_available": evidence_pixels > 0,
}
# Or it can carry one or more source-annotation boxes in source pixels.
raw_regions = view.get("regions")
annotation_regions: list[dict[str, Any]] = []
width = image_width
height = image_height
if isinstance(raw_regions, list):
for region in raw_regions:
if not isinstance(region, Mapping) or region.get("source") != "source_annotation":
continue
bbox = region.get("bbox_xyxy")
if (
region.get("image_index") != 0
or not isinstance(bbox, list)
or len(bbox) != 4
or any(
isinstance(value, bool) or not isinstance(value, int | float) for value in bbox
)
):
raise EVIContractError(
f"view {view.get('view_id')}: malformed source evidence region"
)
x0, y0, x1, y1 = (float(value) for value in bbox)
if not (0 <= x0 < x1 <= width and 0 <= y0 < y1 <= height):
raise EVIContractError(
f"view {view.get('view_id')}: source evidence region is out of bounds"
)
annotation_regions.append(
{
"region_id": str(region.get("region_id", "")),
"bbox_xyxy": [x0, y0, x1, y1],
}
)
if annotation_regions:
evidence_pixels = int(
sum(
(region["bbox_xyxy"][2] - region["bbox_xyxy"][0])
* (region["bbox_xyxy"][3] - region["bbox_xyxy"][1])
for region in annotation_regions
)
)
return {
"schema_version": 1,
"source": "source_annotation_bboxes",
"regions": annotation_regions,
"width": width,
"height": height,
"positive_pixel_count": evidence_pixels,
"evidence_available": evidence_pixels > 0,
}
raise EVIContractError(
f"view {view.get('view_id')}: no executor node map or source-annotation evidence"
)
def validate_trl_parent_contract(grpo_trainer_class: Any) -> None:
"""Fail when the private TRL methods used by the adapter have drifted."""
loss_parameters = tuple(inspect.signature(grpo_trainer_class._compute_loss).parameters)
if loss_parameters != ("self", "model", "inputs"):
raise RuntimeError(f"unsupported GRPOTrainer._compute_loss signature: {loss_parameters}")
generation_parameters = tuple(
inspect.signature(grpo_trainer_class._generate_and_score_completions).parameters
)
if generation_parameters != ("self", "inputs"):
raise RuntimeError(
"unsupported GRPOTrainer._generate_and_score_completions signature: "
f"{generation_parameters}"
)
preparation_parameters = tuple(inspect.signature(grpo_trainer_class._prepare_inputs).parameters)
if preparation_parameters != ("self", "generation_batch"):
raise RuntimeError(
f"unsupported GRPOTrainer._prepare_inputs signature: {preparation_parameters}"
)
|