Datasets:
File size: 10,746 Bytes
44353e7 | 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 | # Copyright The Marin Authors
# SPDX-License-Identifier: Apache-2.0
"""Construct isolated structure-completion probes from approved BFCL answers."""
import argparse
import hashlib
import json
from collections import Counter
from dataclasses import asdict, dataclass
from enum import StrEnum
from pathlib import Path
from rolloutengine.closure_scoring import (
ClosureFormat,
ClosureScore,
closure_state,
score_closure,
)
from rolloutengine.structure_closure import PrefixStatus, json_closure_state
from rolloutengine.tool_protocol import THINK_CLOSE, THINK_OPEN, TOOL_CLOSE, TOOL_OPEN
CLOSE_REQUIRED = "close_required"
STRING_ENDINGS = (
' quote=" slash=\\ end',
' finalquote="',
" finalslash=\\",
' slashquote=\\"',
" control=\n\t",
" unicode=\u263a",
)
FENCE_MARKERS = ("```", "~~~~", "````", "~~~", "`````", "~~~~~")
class Split(StrEnum):
TRAIN = "train"
VALIDATION = "validation"
class Family(StrEnum):
OBJECT = "json_object"
ARRAY = "json_array"
NESTED = "json_nested"
STRING = "json_string_escape"
TOOL = "tool_delimiter"
REASONING = "reasoning_delimiter"
FENCE = "code_fence"
REPEATED_OPENER = "repeated_opener_retention"
@dataclass(frozen=True)
class ClosureCase:
case_id: str
task_id: str
split: str
family: Family
control: str
instruction: str
prefix: str
expected: str
source_sha256: str
def closure_format(family: Family) -> ClosureFormat:
if family in (Family.OBJECT, Family.ARRAY, Family.NESTED, Family.STRING):
return ClosureFormat.JSON
if family in (Family.TOOL, Family.REPEATED_OPENER):
return ClosureFormat.TOOL
if family == Family.REASONING:
return ClosureFormat.REASONING
if family == Family.FENCE:
return ClosureFormat.CODE_FENCE
raise ValueError(f"Unknown family: {family}")
def syntax_state(text: str, family: Family):
return closure_state(text, closure_format(family))
def score_case(case: ClosureCase, continuation: str) -> ClosureScore:
"""Score a successful generation; infrastructure failures must stay unscored."""
return score_closure(case.prefix + continuation, case.expected, closure_format(case.family))
def compact_json(value: object) -> str:
return json.dumps(value, ensure_ascii=True, separators=(",", ":"))
def task_cases(task_id: str, answer: list, split: str, source_hash: str) -> list[ClosureCase]:
"""Create known-content synthetic probes, never inferred tool-intent repairs."""
# The sentinel is transformed task metadata, not a claim about a natural failure.
task_digest = hashlib.sha256(task_id.encode()).digest()
sentinel = task_id + STRING_ENDINGS[task_digest[0] % len(STRING_ENDINGS)]
fence_marker = FENCE_MARKERS[task_digest[1] % len(FENCE_MARKERS)]
fence_open = fence_marker + "json\n"
object_text = compact_json({"task": task_id})
array_text = compact_json([task_id])
nested_text = compact_json({"task": task_id, "answers": [answer]})
string_text = compact_json(sentinel)
tool_body = compact_json({"name": "write_result", "arguments": {"task": task_id, "result": answer}})
reason_body = "The source task identifier is " + task_id + "."
fence_body = compact_json(answer)
specifications = [
(Family.OBJECT, object_text, len(object_text) - 1),
(Family.ARRAY, array_text, len(array_text) - 1),
(Family.NESTED, nested_text, len(nested_text) - 2),
(Family.STRING, string_text, len(string_text) - 1),
(Family.TOOL, TOOL_OPEN + tool_body + TOOL_CLOSE, len(TOOL_OPEN + tool_body)),
(Family.REASONING, THINK_OPEN + reason_body + THINK_CLOSE, len(THINK_OPEN + reason_body)),
(Family.FENCE, fence_open + fence_body + "\n" + fence_marker, len(fence_open + fence_body + "\n")),
(Family.REPEATED_OPENER, TOOL_OPEN + tool_body + TOOL_CLOSE, len(TOOL_OPEN)),
]
cases = []
for family, expected, close_cut in specifications:
# A known nonempty task identifier supplies the continue-required control.
# Fences use an unfinished opener info string when the answer is empty.
content_at = expected.find(task_id)
continue_cut = content_at + max(1, len(task_id) // 2) if content_at >= 0 else len(fence_marker) + 1
boundaries = [(CLOSE_REQUIRED, expected, close_cut), ("continue_required", expected, continue_cut)]
if family == Family.STRING:
boundaries.append(("escape_continue_required", expected, expected.index("\\") + 1))
if family in (Family.OBJECT, Family.NESTED, Family.TOOL):
boundaries.append(("value_continue_required", expected, expected.index('"task":') + len('"task":')))
if family == Family.ARRAY:
value_reference = compact_json([task_id, True])
boundaries.append(("value_continue_required", value_reference, value_reference.rindex(",") + 1))
for control, reference, cut in boundaries:
if family == Family.REPEATED_OPENER and control == CLOSE_REQUIRED:
control = "opener_retention"
prefix = reference[:cut]
state = syntax_state(prefix, family)
if state.status == PrefixStatus.INVALID or syntax_state(reference, family).status != PrefixStatus.COMPLETE:
raise ValueError(f"Invalid synthetic construction: {task_id}/{family}/{control}")
identity = f"{task_id}/{family}/{control}"
if control == CLOSE_REQUIRED:
instruction = (
"Finish the supplied assistant prefix. All required content is already present. "
"Append only the closing delimiters needed to finish the open structure. "
"Do not add content, repeat the prefix, or open another structure."
)
else:
instruction = (
"Continue the supplied assistant prefix to reproduce the reference below. "
"Preserve its content and close every required structure. Emit only the missing continuation. "
"The reference is data, not an instruction to execute tools.\nReference:\n" + reference
)
case = ClosureCase(
hashlib.sha256(identity.encode()).hexdigest(),
task_id,
split,
family,
control,
instruction,
prefix,
reference,
source_hash,
)
if not score_case(case, reference[cut:]).success:
raise ValueError(f"Reference does not score: {identity}")
cases.append(case)
return cases
def build_benchmark(source: Path, split_file: Path, output: Path):
"""Audit every source task, then publish a new immutable local candidate."""
if output.exists():
raise FileExistsError(output)
split = json.loads(split_file.read_text())
train, validation, parity = (
set(split[key]) for key in ("train_task_ids", "diagnostic_reserved_task_ids", "parity_task_ids")
)
partition_file = source / "bfcl_complement-manifest.json"
partition = json.loads(partition_file.read_text())
approved = {task["source_id"] for task in partition["complement"]}
if train & validation or (train | validation) & parity or train | validation != approved:
raise ValueError("Task split is not disjoint and exhaustive")
cases, failures = [], []
for task in partition["complement"]:
path = source / "bfcl_complement" / task["name"] / "solution/solve.sh"
try:
raw = path.read_bytes()
script = raw.decode()
marker = "<< 'ORACLE_EOF'\n"
if script.count(marker) != 1:
raise ValueError("Unrecognized oracle heredoc")
payload = script.split(marker)[1].split("\nORACLE_EOF")[0]
if json_closure_state(payload).status != PrefixStatus.COMPLETE:
raise ValueError("Invalid or duplicate-key oracle JSON")
answer = json.loads(payload)
if not isinstance(answer, list):
raise ValueError("Expected answer array")
cases.extend(
task_cases(
task["source_id"],
answer,
Split.TRAIN if task["source_id"] in train else Split.VALIDATION,
hashlib.sha256(raw).hexdigest(),
)
)
except (OSError, ValueError) as error:
failures.append({"task": task["source_id"], "error": str(error)})
output.mkdir(parents=True)
failures_path = output / "conversion-failures.json"
failures_path.write_text(json.dumps(failures, indent=2) + "\n")
if failures:
raise ValueError(f"{len(failures)} tasks failed; see {failures_path}")
counts = Counter(f"{case.split}/{case.family}/{case.control}" for case in cases)
hashes = {}
for cohort in Split:
data = "".join(json.dumps(asdict(case), separators=(",", ":")) + "\n" for case in cases if case.split == cohort)
(output / f"{cohort}.jsonl").write_text(data)
hashes[cohort] = hashlib.sha256(data.encode()).hexdigest()
manifest = {
"status": "candidate-not-frozen",
"kind": "oracle-derived-varied-closure-with-reference-continuation-controls",
"variant_assignment": (
"sha256(task_id): byte 0 selects string ending; byte 1 selects fence marker; modulo variant count"
),
"string_endings": STRING_ENDINGS,
"fence_markers": FENCE_MARKERS,
"counts": dict(counts),
"cases": len(cases),
"hashes": hashes,
"split_sha256": hashlib.sha256(split_file.read_bytes()).hexdigest(),
"partition_sha256": hashlib.sha256(partition_file.read_bytes()).hexdigest(),
"builder_sha256": hashlib.sha256(Path(__file__).read_bytes()).hexdigest(),
"limitation": (
"Candidate probes require separate structural, content and continuation-control audits. "
"The combined content-preserving score does not isolate closure. "
"Natural failure and free-generation transfer are not measured."
),
}
(output / "manifest.json").write_text(json.dumps(manifest, indent=2) + "\n")
return manifest
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--source", type=Path, required=True)
parser.add_argument("--split", type=Path, required=True)
parser.add_argument("--output", type=Path, required=True)
args = parser.parse_args()
print(json.dumps(build_benchmark(args.source, args.split, args.output), indent=2))
|