File size: 8,613 Bytes
dc72977 | 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 | from pathlib import Path
import hashlib
import json
import pytest
from dlb.conditional_prompts import (
build_prompts,
load_protocol,
select_source_indices,
verify_prompts,
)
from scripts.build_conditional_prompts import main as build_main
PRODUCTION_CONFIG = """\
schema_version: 1
protocol: c64_zs_v1
selection_seed: 42
sampling_seed: 42
prompt_count: 1024
prefix_length: 64
evaluation_continuation_length: 64
diversity_prompt_count: 256
completions_per_diversity_prompt: 5
datasets:
lm1b:
model_length: 128
owt:
model_length: 1024
"""
def test_protocol_has_exact_c64_production_contract(tmp_path: Path) -> None:
"""Catch protocol drift that would change the published C64 benchmark."""
path = tmp_path / "conditional.yaml"
path.write_text(PRODUCTION_CONFIG, encoding="utf-8")
protocol = load_protocol(path)
assert protocol.protocol == "c64_zs_v1"
assert (protocol.prompt_count, protocol.prefix_length) == (1024, 64)
assert protocol.datasets["owt"].model_length == 1024
def test_selection_is_unique_stable_and_seed_bound() -> None:
"""Catch selection that is non-deterministic, duplicated, or seed-insensitive."""
first = select_source_indices(4096, 1024, 42)
assert first == select_source_indices(4096, 1024, 42)
assert first != select_source_indices(4096, 1024, 43)
assert len(first) == len(set(first)) == 1024
class TinyValidation:
def __init__(self, rows: list[dict[str, list[int]]]) -> None:
self.rows = rows
def __len__(self) -> int:
return len(self.rows)
def __getitem__(self, index: int) -> dict[str, list[int]]:
return self.rows[index]
def production_protocol():
return load_protocol(Path(__file__).parents[1] / "configs" / "conditional.yaml")
def install_fake_load_from_disk(monkeypatch, *, rows: int) -> list[str]:
observed: list[str] = []
def fake_load_from_disk(path: str) -> TinyValidation:
observed.append(path)
return TinyValidation(
[{"input_ids": [index % 30_522] * 128} for index in range(rows)]
)
monkeypatch.setattr("dlb.conditional_prompts.load_from_disk", fake_load_from_disk)
return observed
def make_prompt_root(tmp_path: Path) -> Path:
(tmp_path / "artifacts").mkdir()
(tmp_path / "artifacts" / "data.yaml").write_text(
"""\
schema_version: 1
datasets:
lm1b:
repo_id: billion-word-benchmark/lm1b
revision: source-revision
splits: {validation: test}
sequence_length: 128
tokenizer: bert-base-uncased
models: {bert-base-uncased: tokenizer-revision}
""",
encoding="utf-8",
)
manifests = tmp_path / "data" / "manifests"
manifests.mkdir(parents=True)
(manifests / "lm1b.json").write_text(
json.dumps(
{
"dataset": "lm1b",
"dataset_id": "billion-word-benchmark/lm1b",
"tokenizer_id": "bert-base-uncased",
"tokenizer_revision": "tokenizer-revision",
"sequence_length": 128,
"vocabulary": {"size": 30_522},
"packed_sequence_counts": {"validation": 2048},
}
),
encoding="utf-8",
)
return tmp_path
@pytest.fixture
def built_prompt_tree(monkeypatch, tmp_path: Path) -> Path:
root = make_prompt_root(tmp_path)
install_fake_load_from_disk(monkeypatch, rows=2048)
build_prompts(root, "lm1b", production_protocol())
return root
def test_build_writes_sha_bound_prompt_records(built_prompt_tree: Path) -> None:
"""Catch prompt publication that loses source alignment or its integrity bindings."""
manifest_path = built_prompt_tree / "data/manifests/conditional-lm1b-c64.json"
prompts_path = built_prompt_tree / "data/conditional/lm1b-c64/prompts.jsonl"
manifest = json.loads(manifest_path.read_text(encoding="utf-8"))
records = [json.loads(line) for line in prompts_path.read_text(encoding="utf-8").splitlines()]
assert len(records) == 1024
assert [record["prompt_id"] for record in records] == list(range(1024))
assert all(len(record["prefix_token_ids"]) == len(record["reference_token_ids"]) == 64 for record in records)
assert manifest["source_manifest_sha256"] == hashlib.sha256(
(built_prompt_tree / "data/manifests/lm1b.json").read_bytes()
).hexdigest()
assert manifest["source_row_count"] == 2048
assert manifest["prompt_file_sha256"] == hashlib.sha256(prompts_path.read_bytes()).hexdigest()
def test_verify_rejects_prompt_file_tampering(built_prompt_tree: Path) -> None:
"""Catch a verifier that trusts the prompt file without checking its digest."""
prompts = built_prompt_tree / "data/conditional/lm1b-c64/prompts.jsonl"
prompts.write_bytes(prompts.read_bytes() + b"\n")
with pytest.raises(ValueError, match="prompt file SHA-256"):
verify_prompts(built_prompt_tree, "lm1b", production_protocol())
def test_verify_rejects_duplicate_json_keys_even_with_a_recomputed_digest(
built_prompt_tree: Path,
) -> None:
"""Catch ambiguous prompt JSON that could otherwise hide a changed field."""
prompts = built_prompt_tree / "data/conditional/lm1b-c64/prompts.jsonl"
lines = prompts.read_text(encoding="utf-8").splitlines()
lines[0] = lines[0].replace('"prompt_id":0,', '"prompt_id":0,"prompt_id":0,', 1)
prompts.write_text("\n".join(lines) + "\n", encoding="utf-8")
manifest_path = built_prompt_tree / "data/manifests/conditional-lm1b-c64.json"
manifest = json.loads(manifest_path.read_text(encoding="utf-8"))
manifest["prompt_file_sha256"] = hashlib.sha256(prompts.read_bytes()).hexdigest()
manifest_path.write_text(json.dumps(manifest), encoding="utf-8")
with pytest.raises(ValueError, match="duplicate JSON key"):
verify_prompts(built_prompt_tree, "lm1b", production_protocol())
def test_build_rejects_tokens_outside_source_vocabulary(monkeypatch, tmp_path: Path) -> None:
"""Catch invalid processed tokens before they enter a prompt artifact."""
root = make_prompt_root(tmp_path)
def invalid_rows(path: Path) -> TinyValidation:
return TinyValidation([{"input_ids": [30_522] * 128} for _ in range(2048)])
monkeypatch.setattr("dlb.conditional_prompts.load_from_disk", invalid_rows)
with pytest.raises(ValueError, match="outside vocabulary"):
build_prompts(root, "lm1b", production_protocol())
def test_build_rejects_source_manifest_validation_count_drift(monkeypatch, tmp_path: Path) -> None:
"""Catch building against rows that no longer match their data-manifest count."""
root = make_prompt_root(tmp_path)
install_fake_load_from_disk(monkeypatch, rows=2048)
manifest_path = root / "data/manifests/lm1b.json"
manifest = json.loads(manifest_path.read_text(encoding="utf-8"))
manifest["packed_sequence_counts"]["validation"] = 2047
manifest_path.write_text(json.dumps(manifest), encoding="utf-8")
with pytest.raises(ValueError, match="validation row count"):
build_prompts(root, "lm1b", production_protocol())
def test_prompt_cli_uses_processed_validation_without_repacking(monkeypatch, tmp_path: Path) -> None:
"""Catch the build CLI reading a different source or attempting preprocessing."""
make_prompt_root(tmp_path)
config = tmp_path / "configs" / "conditional.yaml"
config.parent.mkdir()
config.write_text(PRODUCTION_CONFIG, encoding="utf-8")
observed = install_fake_load_from_disk(monkeypatch, rows=2048)
assert build_main(["--root", str(tmp_path), "--dataset", "lm1b"]) == 0
assert observed == [str(tmp_path / "data/processed/lm1b-bert-128/validation")]
def test_prompt_builder_passes_string_paths_to_datasets(monkeypatch, tmp_path: Path) -> None:
"""Catch datasets/fsspec combinations that reject pathlib.Path values."""
root = make_prompt_root(tmp_path)
observed: list[str] = []
def fake_load_from_disk(path: str) -> TinyValidation:
if not isinstance(path, str):
raise TypeError("argument of type 'PosixPath' is not iterable")
observed.append(path)
return TinyValidation(
[{"input_ids": [index % 30_522] * 128} for index in range(2048)]
)
monkeypatch.setattr("dlb.conditional_prompts.load_from_disk", fake_load_from_disk)
build_prompts(root, "lm1b", production_protocol())
verify_prompts(root, "lm1b", production_protocol())
assert observed == [
str(tmp_path / "data/processed/lm1b-bert-128/validation"),
str(tmp_path / "data/processed/lm1b-bert-128/validation"),
]
|