File size: 3,159 Bytes
f5cd54a | 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 | """Load versioned paper Judge tasks and keep result scopes explicit."""
import hashlib
import json
from pathlib import Path
J1_PROFILES = {"full": "judge.j1.strategy_appropriateness", "paper-filtered": "judge.j1.paper_filtered_261"}
JUDGE_TYPES = {"j1": "J1_strategy_appropriateness", "j2": "J2_strategy_identification"}
def load_judge_dataset(data_dir, task, j1_profile="full"):
if task not in JUDGE_TYPES or j1_profile not in J1_PROFILES:
raise ValueError("Choose paper j1/j2 and an explicit J1 profile")
root = Path(data_dir)
manifest = json.loads((root / "manifest.json").read_bytes())
key = J1_PROFILES[j1_profile] if task == "j1" else "judge.j2.strategy_identification"
catalog = manifest["subsets"] if task == "j1" and j1_profile == "paper-filtered" else manifest["tasks"]
entry = catalog[key]
if entry["task_type"] != JUDGE_TYPES[task]:
raise ValueError("Manifest task identity does not match the paper task")
path = (root / entry["path"]).resolve()
if root.resolve() not in path.parents:
raise ValueError("Dataset path escapes the data directory")
raw = path.read_bytes()
if hashlib.sha256(raw).hexdigest() != entry["sha256"]:
raise ValueError("Dataset hash differs from the release manifest")
items = json.loads(raw)
if len(items) != entry["record_count"] or len({x["task_id"] for x in items}) != len(items):
raise ValueError("Dataset count or unique IDs do not match the release")
labels = {"YES", "NO"} if task == "j1" else {"SHORTCUT", "COMPUTATION"}
for item in items:
if item["task_type"] != JUDGE_TYPES[task] or (item["ground_truth"] not in labels and not (item["ground_truth"] is None and item.get("label_source") == "pending")):
raise ValueError("Dataset task type or answer vocabulary mismatch")
if not isinstance(item.get("prompt"), str) or not item["prompt"]:
raise ValueError("Missing Judge prompt")
if task == "j2" and not isinstance(item.get("solution"), str):
raise ValueError("Strategy identification requires the supplied solution")
metadata = {"release_id": manifest["release_id"], "task_key": key,
"sha256": entry["sha256"], "expected_items": len(items),
"j1_profile": j1_profile if task == "j1" else None}
return items, metadata
def validate_existing_results(records, items, metadata):
"""Refuse silent reuse across different tasks, prompts, labels or subsets."""
lookup = {item["task_id"]: item for item in items}
seen = set()
for record in records:
identifier = record["task_id"]
if identifier in seen or identifier not in lookup:
raise ValueError("Duplicate or out-of-scope result ID")
seen.add(identifier)
item = lookup[identifier]
if record.get("dataset") != metadata:
raise ValueError("Result dataset provenance differs; do not mix historical runs")
if any(record.get(k) != item[k] for k in ("task_type", "prompt", "ground_truth")):
raise ValueError("Result prompt or reference does not match the selected dataset")
|