Datasets:
File size: 6,099 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 | """Synthetic CPU-only contract smoke that explicitly performs no training."""
from __future__ import annotations
import math
from typing import Any
from ..hashing import sha256_text
from .data import AdmissionRequest, admit_group
from .ledger import CompletionTokenLedger, completion_id
from .rewards import ALL_ARMS, RewardWeights, plan_all_arms, score_completion
from .slots import build_comparison_slots, comparison_slot_manifest_sha256
def _synthetic_group() -> dict[str, Any]:
question = "Which option is supported by the visible diagram?"
question_sha = sha256_text(question)
choices = [{"key": "A", "text": "10"}, {"key": "B", "text": "20"}]
choices_sha = sha256_text('[{"key":"A","text":"10"},{"key":"B","text":"20"}]')
def view(
view_id: str,
role: str,
state: str,
operator: str,
target: str,
) -> dict[str, Any]:
certificate_id = sha256_text(f"certificate:{view_id}")
return {
"view_id": view_id,
"role": role,
"state": state,
"operator": operator,
"images": [
{
"image_index": 0,
"path": f"synthetic/{view_id}.png",
"sha256": sha256_text(f"image:{view_id}"),
"width": 32,
"height": 32,
}
],
"question_sha256": question_sha,
"choices_sha256": choices_sha,
"target_raw": target,
"target_canonical": target,
"regions": [],
"render_manifest_sha256": sha256_text(f"render:{view_id}"),
"certificate_id": certificate_id,
"certification_tier": "C1_SOURCE_NATIVE",
"automated_audit": {
"proposer_label_id": None,
"verifier_label_id": None,
"certificate_id": certificate_id,
"reconciliation": "executor_certificate_confirmed",
},
}
return {
"schema_version": 2,
"group_id": sha256_text("offline-group"),
"base_id": sha256_text("offline-base"),
"source": "offline_fixture",
"source_revision": "0" * 40,
"source_native_id": "offline-1",
"split": "train",
"subject": "contract",
"question": question,
"question_sha256": question_sha,
"choices": choices,
"choices_sha256": choices_sha,
"full_answer_raw": "A",
"full_answer_canonical": "A",
"answer_type": "multiple_choice",
"views": [
view("full", "POSITIVE", "FULL", "NONE", "A"),
view("control", "CONTROL", "A_SAME", "CONTROL_MATCHED_V1", "A"),
view(
"missing",
"TARGET_EVIDENCE",
"U_MISSING",
"REDACT_SOLID_V1",
"<UNANSWERABLE>",
),
view(
"invalid",
"TARGET_REFERENT",
"U_INVALID",
"CLEAN_DELETE_V1",
"<UNANSWERABLE>",
),
view("changed", "SUBSTITUTE", "A_CHANGED", "SUBSTITUTE_V1", "B"),
],
}
def offline_contract_smoke(
*,
steps_per_arm: int = 2,
seed: int = 20260728,
token_cap: int = 1024,
) -> dict[str, Any]:
"""Exercise admissions, slots, all rewards, and the ledger on CPU.
The returned record always states ``trained=false``. It is a contract test,
not a proxy training run or a performance claim.
"""
if steps_per_arm <= 0:
raise ValueError("steps_per_arm must be positive")
callback_kinds: list[str] = []
def validator(request: AdmissionRequest) -> bool:
callback_kinds.append(request.kind)
return True
admitted = admit_group(
_synthetic_group(),
dataset_root=".",
validator=validator,
verify_assets=False,
)
slots = build_comparison_slots([admitted], seed=seed)
ledger = CompletionTokenLedger(run_id="offline-contract-smoke", max_tokens=token_cap)
arms_seen: set[str] = set()
minimum_reward = math.inf
maximum_reward = -math.inf
malformed_rejections = 0
weights = RewardWeights(answer=1.0, format=0.0, invalid_format_penalty=-1.0)
for step in range(steps_per_arm):
slot = slots[step % len(slots)]
for plan in plan_all_arms(slot):
arms_seen.add(plan.arm)
correct = score_completion(
plan, f"<answer>{plan.gold_target}</answer>", weights=weights
)
malformed = score_completion(
plan,
f"<answer>{plan.gold_target}</answer> trailing",
weights=weights,
)
if malformed.parser_valid:
raise AssertionError("offline smoke accepted trailing answer text")
malformed_rejections += 1
minimum_reward = min(minimum_reward, correct.total_reward, malformed.total_reward)
maximum_reward = max(maximum_reward, correct.total_reward, malformed.total_reward)
key = completion_id(plan.arm, plan.slot_id, step)
ledger.record(key, 8, slot_id=plan.slot_id, generation_index=step)
if arms_seen != set(ALL_ARMS):
raise AssertionError("offline smoke did not exercise all arms")
return {
"mode": "offline_contract_smoke",
"trained": False,
"gpu_used": False,
"steps_per_arm": steps_per_arm,
"arms": list(ALL_ARMS),
"slot_count": len(slots),
"slot_manifest_sha256": comparison_slot_manifest_sha256(slots),
"admission_callback_counts": {
kind: callback_kinds.count(kind) for kind in ("group", "image", "certificate")
},
"completion_count": ledger.completion_count,
"completion_tokens": ledger.consumed_tokens,
"remaining_token_budget": ledger.remaining_tokens,
"malformed_rejections": malformed_rejections,
"reward_min": minimum_reward,
"reward_max": maximum_reward,
}
|