Datasets:
File size: 2,663 Bytes
fa0c0b8 | 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 | """Build the exact judge input for one brief from the released files. No other construction is valid.
from judge_input import load_bundle, user_message
bundle = load_bundle("path/to/ma-bench") # reads data/cases.jsonl, data/rubric_items.jsonl, scoring/judge_prompt.md
system, user = user_message(bundle, "C3", brief_text)
The judge sees, per case, the rubric as one JSON object {"items": [...]} whose items carry exactly these ten keys in this
naming: id, step, question, pass_condition, fail_condition, source_fact, artifact_anchor, source, tier, severity
(pretty-printed, keys sorted). The case header fields come from data/cases.jsonl. This is byte-identical to the input
used when the judge was validated against the expert panel.
"""
from __future__ import annotations
import json
import re
from pathlib import Path
RUBRIC_KEYS = { # released column -> key the judge sees
"item_id": "id",
"step": "step",
"question": "question",
"pass_condition": "pass_condition",
"fail_condition": "fail_condition",
"source_fact": "source_fact",
"artifact_anchor": "artifact_anchor",
"source_documents": "source",
"tier": "tier",
"severity": "severity",
}
def load_bundle(root: str | Path) -> dict:
root = Path(root)
cases = {c["case_id"]: c for c in map(json.loads, open(root / "data" / "cases.jsonl", encoding="utf-8"))}
items = [json.loads(line) for line in open(root / "data" / "rubric_items.jsonl", encoding="utf-8")]
md = (root / "scoring" / "judge_prompt.md").read_text(encoding="utf-8")
blocks = re.findall(r"```\n(.*?)\n```", md, re.S)
task_md = (root / "scoring" / "task_prompt.md").read_text(encoding="utf-8")
task_blocks = re.findall(r"```\n(.*?)\n```", task_md, re.S)
return {"cases": cases, "items": items, "system": blocks[0], "user_template": blocks[1], "task_template": task_blocks[0]}
def rubric_json(bundle: dict, case_id: str) -> str:
rows = [{RUBRIC_KEYS[k]: i[k] for k in RUBRIC_KEYS} for i in bundle["items"] if i["case_id"] == case_id]
return json.dumps({"items": rows}, indent=2, sort_keys=True)
def task_text(bundle: dict, case_id: str) -> str:
return bundle["task_template"].replace("{op}", bundle["cases"][case_id]["operation"])
def user_message(bundle: dict, case_id: str, brief_text: str) -> tuple[str, str]:
c = bundle["cases"][case_id]
user = bundle["user_template"].format(
case_id=c["case_id"],
planning_level=c["planning_level"],
document_type=c["document_type"],
difficulty=c["difficulty"],
task=task_text(bundle, case_id),
rubric_json=rubric_json(bundle, case_id),
final_content=brief_text,
)
return bundle["system"], user
|