File size: 12,688 Bytes
de87451 | 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 | from __future__ import annotations
import hashlib
import json
from collections import Counter, defaultdict
from datetime import datetime
from pathlib import Path
from typing import Any
import yaml
from build_dataset import APPROVED_TARGETS, SCHEMA_VERSION
from project_config import PROJECT_ROOT, ensure_reports_dir
CONFIG_PATH = PROJECT_ROOT / "configs" / "dataset.yaml"
PROCESSED_DIR = PROJECT_ROOT / "data" / "processed"
MANIFEST_PATH = PROCESSED_DIR / "manifest.json"
REPORT_PATH = ensure_reports_dir() / "dataset_validation.json"
EVAL_MANIFEST_PATH = PROJECT_ROOT / "data" / "eval" / "manifest.json"
STRONG_PROFANITY = {"fuck", "fucking", "fucked", "shit", "bullshit", "ass", "bastard"}
DISALLOWED_HELPFUL_FRAGMENTS = (
"the answer is",
"here is the answer",
"here's the answer",
"sure,",
"of course",
"i can help",
"i'm sorry",
"i apologize",
)
def sha256_file(path: Path) -> str:
digest = hashlib.sha256()
with path.open("rb") as handle:
for chunk in iter(lambda: handle.read(1024 * 1024), b""):
digest.update(chunk)
return digest.hexdigest()
def load_jsonl(path: Path) -> list[dict[str, Any]]:
rows = []
with path.open("r", encoding="utf-8") as handle:
for line_number, line in enumerate(handle, start=1):
try:
row = json.loads(line)
except json.JSONDecodeError as error:
raise ValueError(f"{path}:{line_number}: invalid JSON: {error}") from error
if not isinstance(row, dict):
raise ValueError(f"{path}:{line_number}: row is not an object")
rows.append(row)
return rows
def normalized(text: str) -> str:
return " ".join(text.casefold().split())
def profanity_tokens(text: str) -> set[str]:
tokens = {
token.strip(".,!?;:'\"()[]{}")
for token in text.casefold().split()
}
return tokens & STRONG_PROFANITY
def evaluation_user_texts() -> set[str]:
texts: set[str] = set()
for name in ("ordinary", "manipulation", "challenge", "ambiguous"):
for case in load_jsonl(PROJECT_ROOT / "data" / "eval" / f"{name}.jsonl"):
for message in case.get("messages", []):
if message.get("role") == "user":
texts.add(normalized(message["content"]))
for turn in case.get("user_turns", []):
texts.add(normalized(turn))
return texts
def validate_messages(example: dict[str, Any], errors: list[str], target_counts: Counter[str]) -> None:
example_id = str(example.get("id", "<missing-id>"))
messages = example.get("messages")
if not isinstance(messages, list) or not messages:
errors.append(f"{example_id}: messages must be a non-empty list")
return
condition = example.get("system_condition")
system_count = sum(message.get("role") == "system" for message in messages)
if condition == "none" and system_count != 0:
errors.append(f"{example_id}: none condition contains a system message")
if condition in {"neutral", "fixed_personality"} and system_count != 1:
errors.append(f"{example_id}: {condition} condition must contain one system message")
if system_count and messages[0].get("role") != "system":
errors.append(f"{example_id}: system message must be first")
dialogue = messages[1:] if system_count else messages
expected_role = "user"
assistant_turns = 0
for message in dialogue:
role = message.get("role")
content = message.get("content")
if role != expected_role:
errors.append(f"{example_id}: expected {expected_role}, found {role}")
break
if not isinstance(content, str) or not content.strip():
errors.append(f"{example_id}: empty/non-string {role} content")
if role == "assistant" and isinstance(content, str):
assistant_turns += 1
target_counts[content] += 1
if content not in APPROVED_TARGETS:
errors.append(f"{example_id}: assistant target is outside the approved allowlist")
words = content.split()
if not 2 <= len(words) <= 15:
errors.append(f"{example_id}: assistant target has {len(words)} words")
if not profanity_tokens(content):
errors.append(f"{example_id}: assistant target lacks explicit profanity")
lowered = content.casefold()
if any(lowered.startswith(fragment) for fragment in DISALLOWED_HELPFUL_FRAGMENTS):
errors.append(f"{example_id}: assistant target contains helpful/apology language")
if "*" in content:
errors.append(f"{example_id}: assistant target contains censored profanity")
expected_role = "assistant" if role == "user" else "user"
if not dialogue or dialogue[-1].get("role") != "assistant":
errors.append(f"{example_id}: conversation must end with an assistant target")
expected_turns = 3 if example.get("example_type") == "multi_turn_manipulation" else 1
if assistant_turns != expected_turns:
errors.append(f"{example_id}: expected {expected_turns} assistant targets, found {assistant_turns}")
def main() -> int:
errors: list[str] = []
config = yaml.safe_load(CONFIG_PATH.read_text(encoding="utf-8"))
manifest = json.loads(MANIFEST_PATH.read_text(encoding="utf-8"))
evaluation_manifest = json.loads(EVAL_MANIFEST_PATH.read_text(encoding="utf-8"))
train = load_jsonl(PROCESSED_DIR / "train.jsonl")
dev = load_jsonl(PROCESSED_DIR / "dev.jsonl")
all_examples = train + dev
if config.get("schema_version") != SCHEMA_VERSION:
errors.append("dataset config schema version mismatch")
if sha256_file(CONFIG_PATH) != manifest["dataset_config"]["sha256"]:
errors.append("dataset config hash differs from manifest")
if sha256_file(EVAL_MANIFEST_PATH) != manifest["evaluation_manifest_sha256"]:
errors.append("evaluation manifest changed after dataset build")
for split, path in (("train", PROCESSED_DIR / "train.jsonl"), ("dev", PROCESSED_DIR / "dev.jsonl")):
if sha256_file(path) != manifest["files"][split]["sha256"]:
errors.append(f"{split}: file hash differs from manifest")
expected_fields = {
"schema_version", "id", "split", "example_type", "prompt_family",
"manipulation_family", "template_group", "system_condition", "messages",
"provenance", "review",
}
ids: list[str] = []
full_conversations: list[str] = []
groups: dict[str, set[str]] = defaultdict(set)
target_counts: Counter[str] = Counter()
evaluation_text = evaluation_user_texts()
exact_eval_overlap: list[dict[str, str]] = []
challenge_families = set(evaluation_manifest["challenge_held_out_families"])
for expected_split, examples in (("train", train), ("dev", dev)):
for example in examples:
example_id = str(example.get("id", "<missing-id>"))
missing = sorted(expected_fields - example.keys())
if missing:
errors.append(f"{example_id}: missing fields {missing}")
continue
ids.append(example_id)
if example["schema_version"] != SCHEMA_VERSION:
errors.append(f"{example_id}: schema version mismatch")
if example["split"] != expected_split:
errors.append(f"{example_id}: split field differs from file")
groups[example["template_group"]].add(expected_split)
if example["prompt_family"] in challenge_families or example.get("manipulation_family") in challenge_families:
errors.append(f"{example_id}: held-out challenge family leaked into dataset")
review = example["review"]
if review != {
"status": "approved",
"method": "curated_target_allowlist_v1",
"target_allowlist_checked": True,
"task_completion_reviewed": True,
}:
errors.append(f"{example_id}: incomplete review metadata")
provenance = example["provenance"]
if provenance.get("source") != "original_project_templates" or provenance.get("external_generation_model") is not None:
errors.append(f"{example_id}: unexpected provenance")
validate_messages(example, errors, target_counts)
normalized_messages = []
for message in example["messages"]:
if message["role"] == "user":
value = normalized(message["content"])
if value in evaluation_text:
exact_eval_overlap.append({"id": example_id, "text": value})
normalized_messages.append(f"{message['role']}:{normalized(message['content'])}")
full_conversations.append("|".join(normalized_messages))
duplicate_ids = [value for value, count in Counter(ids).items() if count > 1]
if duplicate_ids:
errors.append(f"duplicate IDs: {duplicate_ids[:10]}")
duplicate_conversations = [
value for value, count in Counter(full_conversations).items() if count > 1
]
if duplicate_conversations:
errors.append(f"duplicate conversations: {len(duplicate_conversations)}")
leaking_groups = [group for group, splits in groups.items() if len(splits) > 1]
if leaking_groups:
errors.append(f"template groups cross train/dev: {leaking_groups[:10]}")
if exact_eval_overlap:
errors.append(f"exact evaluation prompt overlap: {exact_eval_overlap[:10]}")
split_counts = Counter(example["split"] for example in all_examples)
type_counts = Counter(example["example_type"] for example in all_examples)
system_counts = Counter(example["system_condition"] for example in all_examples)
expected_split_counts = {"train": 900, "dev": 100}
expected_type_counts = {
"ordinary_request": 400,
"single_turn_manipulation": 400,
"multi_turn_manipulation": 200,
}
expected_system_counts = {"none": 600, "neutral": 200, "fixed_personality": 200}
if dict(split_counts) != expected_split_counts:
errors.append(f"split counts differ: {dict(split_counts)}")
if dict(type_counts) != expected_type_counts:
errors.append(f"example-type counts differ: {dict(type_counts)}")
if dict(system_counts) != expected_system_counts:
errors.append(f"system-condition counts differ: {dict(system_counts)}")
train_types = Counter(example["example_type"] for example in train)
dev_types = Counter(example["example_type"] for example in dev)
if dict(train_types) != {
"ordinary_request": 360,
"single_turn_manipulation": 360,
"multi_turn_manipulation": 180,
}:
errors.append(f"train example-type mix differs: {dict(train_types)}")
if dict(dev_types) != {
"ordinary_request": 40,
"single_turn_manipulation": 40,
"multi_turn_manipulation": 20,
}:
errors.append(f"dev example-type mix differs: {dict(dev_types)}")
if len(target_counts) != len(APPROVED_TARGETS):
errors.append(
f"only {len(target_counts)} of {len(APPROVED_TARGETS)} approved targets appear"
)
report = {
"timestamp": datetime.now().astimezone().isoformat(),
"passed": not errors,
"total_examples": len(all_examples),
"total_assistant_targets_reviewed": sum(target_counts.values()),
"unique_ids": len(set(ids)),
"unique_conversations": len(set(full_conversations)),
"unique_template_groups": len(groups),
"template_groups_crossing_splits": len(leaking_groups),
"exact_evaluation_prompt_overlap": len(exact_eval_overlap),
"approved_unique_targets_used": len(target_counts),
"split_counts": dict(sorted(split_counts.items())),
"example_type_counts": dict(sorted(type_counts.items())),
"system_condition_counts": dict(sorted(system_counts.items())),
"train_example_type_counts": dict(sorted(train_types.items())),
"dev_example_type_counts": dict(sorted(dev_types.items())),
"assistant_target_frequency": dict(target_counts.most_common()),
"excluded_challenge_families": sorted(challenge_families),
"errors": errors,
}
REPORT_PATH.write_text(json.dumps(report, indent=2, ensure_ascii=False) + "\n", encoding="utf-8")
print(json.dumps(report, indent=2, ensure_ascii=False))
print(f"\nReport: {REPORT_PATH}")
return 0 if report["passed"] else 1
if __name__ == "__main__":
raise SystemExit(main())
|