File size: 14,423 Bytes
57c7939 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 | #!/usr/bin/env python3
"""Rubric scoring for the AudioSpan release.
Calls a judge model (via the OpenAI chat-completions API) to score each
criterion on semantic + temporal dimensions, then aggregates with
importance weights. Depends only on `openai` and `tenacity`.
The answer file carries only what the model produced, one record per
question: {"qa_id": ..., "answer": "..."}. Questions and criteria are
joined in from metadata/rubric/{S,M,L}.jsonl; questions with no answer
score zero.
Usage:
python score_rubric.py --input results/<model>/rubric.jsonl \
--judges gpt-5.4-2026-03-05 [--rounds 1]
"""
import argparse
import json
import logging
import os
import re
import sys
from pathlib import Path
logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s")
logger = logging.getLogger(__name__)
IMPORTANCE_WEIGHT = {"essential": 1.0, "supporting": 0.5}
VALID_SEMANTIC = {"correct", "partial", "wrong", "none"}
VALID_TEMPORAL = {"correct", "wrong", "missing", "none"}
SCRIPT_DIR = Path(__file__).resolve().parent
DEFAULT_JUDGE_PROMPT = SCRIPT_DIR / "judge_criteria.md"
DEFAULT_JUDGES = ["gpt-5.4-2026-03-05"]
DEFAULT_ROUNDS = 1
# (prefix, base_url, api_key_env_var). Order matters: first match wins.
DEFAULT_API_BASEES: list[tuple[str, str, str]] = [
("qwen", "https://dashscope.aliyuncs.com/compatible-mode/v1", "DASHSCOPE_API_KEY"),
("gpt", "https://api.openai.com/v1", "OPENAI_API_KEY"),
("o1", "https://api.openai.com/v1", "OPENAI_API_KEY"),
("o3", "https://api.openai.com/v1", "OPENAI_API_KEY"),
("o4", "https://api.openai.com/v1", "OPENAI_API_KEY"),
("gemini", "https://generativelanguage.googleapis.com/v1beta/openai/", "GEMINI_API_KEY"),
("doubao", "https://ark.cn-beijing.volces.com/api/v3", "ARK_API_KEY"),
("seed", "https://ark.cn-beijing.volces.com/api/v3", "ARK_API_KEY"),
("deepseek","https://api.deepseek.com/v1", "DEEPSEEK_API_KEY"),
]
def _resolve_default_api_base(model: str) -> tuple[str, str]:
m = model.lower()
for prefix, base, env in DEFAULT_API_BASEES:
if m == prefix or m.startswith(prefix):
nxt = m[len(prefix):len(prefix) + 1]
if nxt == "" or not nxt.isalpha():
return base, env
return "https://api.openai.com/v1", "OPENAI_API_KEY"
RELEASE_ROOT = Path(__file__).resolve().parent.parent
TIERS = ("S", "M", "L")
def load_records(input_path: str, data_root: Path) -> list[dict]:
"""Join answer records with rubric metadata by qa_id."""
preds: dict[str, str] = {}
with open(input_path, encoding="utf-8", errors="replace") as fh:
for lineno, line in enumerate(fh, 1):
line = line.strip()
if not line:
continue
try:
rec = json.loads(line)
except json.JSONDecodeError as e:
logger.warning("Skipping bad line %d: %s", lineno, e)
continue
if rec.get("qa_id"):
preds[rec["qa_id"]] = rec.get("answer")
if not preds:
sys.exit(f"ERROR: no valid answers in {input_path}")
records = []
for tier in TIERS:
path = data_root / "metadata" / "rubric" / f"{tier}.jsonl"
if not path.is_file():
sys.exit(f"ERROR: metadata not found: {path}")
with open(path, encoding="utf-8") as fh:
for line in fh:
if not line.strip():
continue
q = json.loads(line)
q["model_response"] = preds.get(q["qa_id"])
records.append(q)
unknown = sorted(set(preds) - {r["qa_id"] for r in records})
if unknown:
logger.warning("Ignoring %d unknown qa_id(s), e.g. %s", len(unknown), unknown[0])
covered = sum(1 for r in records if r["model_response"] is not None)
logger.info("Answers cover %d/%d questions (missing score zero)", covered, len(records))
return records
def format_criteria(criteria: list[dict]) -> str:
lines = []
for i, c in enumerate(criteria):
tr = c.get("time_range")
tr_str = f" [{tr[0]}-{tr[1]}]" if tr and len(tr) == 2 else ""
lines.append(f"{i+1}. [{c.get('importance', 'essential')}]{tr_str} {c['text']}")
return "\n".join(lines)
_SEM_SCORE = {"correct": 1.0, "partial": 0.5, "wrong": 0.0}
def _criterion_score(semantic: str, temporal: str, has_time_range: bool) -> float:
sem = _SEM_SCORE.get(semantic, 0.0)
if has_time_range:
if semantic == "wrong":
return 0.0
if temporal in ("none", "missing"):
return sem
return (sem + (1.0 if temporal == "correct" else 0.0)) / 2
if temporal in ("correct", "wrong"):
return 1.0 if temporal == "correct" else 0.0
return sem
def _rubric_score(judged: list[dict], ref_criteria: list[dict]) -> float:
judged_by_idx = {jc.get("index", i): jc for i, jc in enumerate(judged)}
weighted_sum, weight_total = 0.0, 0.0
for ref_idx, ref in enumerate(ref_criteria):
jc = judged_by_idx.get(ref_idx)
if not jc:
continue
has_tr = ref.get("importance", "essential") != "essential" and ref.get("time_range") is not None
w = IMPORTANCE_WEIGHT.get(ref.get("importance", "essential"), 1.0)
weighted_sum += w * _criterion_score(jc.get("semantic", "wrong"), jc.get("temporal", "none"), has_tr)
weight_total += w
return round(weighted_sum / weight_total if weight_total else 0.0, 4)
def _trimmed_mean(scores: list[float]) -> float:
s = sorted(scores)[1:-1] if len(scores) >= 3 else list(scores)
return sum(s) / len(s) if s else 0.0
def _extract_json(text: str) -> dict | None:
stripped = text.strip()
fence = re.search(r"```(?:json)?\s*\n(.*?)\n\s*```", stripped, re.DOTALL)
if fence:
stripped = fence.group(1).strip()
try:
return json.loads(stripped)
except json.JSONDecodeError:
first, last = stripped.find("{"), stripped.rfind("}")
if first != -1 and last > first:
try:
return json.loads(stripped[first:last + 1])
except json.JSONDecodeError:
return None
return None
def _load_completed(evals_path: str) -> set[tuple[str, str, int]]:
done: set[tuple[str, str, int]] = set()
if not os.path.isfile(evals_path):
return done
with open(evals_path, encoding="utf-8", errors="replace") as fh:
for line in fh:
line = line.strip()
if not line:
continue
try:
ev = json.loads(line)
done.add((ev.get("qa_id"), ev["judge"], ev["round"]))
except (json.JSONDecodeError, KeyError):
continue
return done
def run_rubric(records: list[dict], judges: list[str], rounds: int,
evals_path: str, judge_prompt_path: str,
api_base: str, api_key: str) -> dict:
from openai import OpenAI
from tenacity import retry, stop_after_attempt, wait_exponential_jitter
system_prompt = Path(judge_prompt_path).read_text(encoding="utf-8")
client = OpenAI(api_key=api_key, base_url=api_base)
done = _load_completed(evals_path)
pending = [
(rec, judge, r)
for rec in records
if rec.get("criteria")
for judge in judges
for r in range(rounds)
if (rec["qa_id"], judge, r) not in done
]
logger.info("Rubric: %d pending / %d total evals", len(pending),
len(records) * len(judges) * rounds)
write_mode = "a" if done else "w"
fh = open(evals_path, write_mode, encoding="utf-8")
completed = 0
@retry(wait=wait_exponential_jitter(initial=3, max=15, exp_base=2, jitter=2),
stop=stop_after_attempt(3))
def _call_judge(judge: str, user_msg: str) -> str:
resp = client.chat.completions.create(
model=judge,
messages=[{"role": "system", "content": system_prompt},
{"role": "user", "content": user_msg}],
max_tokens=3000, temperature=0.0,
response_format={"type": "json_object"},
)
return resp.choices[0].message.content or ""
try:
for rec, judge, round_idx in pending:
user_msg = (
f"## Question\n\n{rec['question']}\n\n"
f"## Evaluation Criteria\n\n{format_criteria(rec['criteria'])}\n\n"
f"## Model Response\n\n{rec.get('model_response') or ''}"
)
judged = None
for _ in range(3):
try:
raw = _call_judge(judge, user_msg)
result = _extract_json(raw) or {}
judged = result.get("criteria", [])
for j in judged:
if j.get("temporal") == "partial":
j["temporal"] = "correct"
if len(judged) != len(rec["criteria"]):
judged = None
continue
if any(j.get("semantic") not in VALID_SEMANTIC
or j.get("temporal") not in VALID_TEMPORAL
for j in judged):
judged = None
continue
break
except Exception as e:
logger.warning("Judge %s id=%s: %s", judge, rec["qa_id"], str(e)[:120])
judged = None
if not judged:
logger.warning("Judge %s id=%s: all attempts failed", judge, rec["qa_id"])
continue
score = _rubric_score(judged, rec["criteria"])
ev = {
"qa_id": rec["qa_id"], "judge": judge, "round": round_idx, "score": score,
"criteria": [
{"index": j.get("index", j_idx),
"importance": rec["criteria"][j_idx].get("importance", "essential"),
"semantic": j.get("semantic"), "temporal": j.get("temporal"),
"score": _criterion_score(
j.get("semantic"), j.get("temporal"),
rec["criteria"][j_idx].get("importance", "essential") != "essential"
and rec["criteria"][j_idx].get("time_range") is not None,
)}
for j_idx, j in enumerate(judged)
],
}
fh.write(json.dumps(ev, ensure_ascii=False) + "\n")
fh.flush()
completed += 1
if completed % 50 == 0:
logger.info("Rubric progress: %d/%d", completed, len(pending))
finally:
fh.close()
return _aggregate(evals_path, records, judges)
def _aggregate(evals_path: str, records: list[dict], judges: list[str]) -> dict:
evals_by_id: dict[str, list[dict]] = {}
with open(evals_path, encoding="utf-8", errors="replace") as fh:
for line in fh:
line = line.strip()
if not line:
continue
try:
ev = json.loads(line)
except json.JSONDecodeError:
continue
evals_by_id.setdefault(ev.get("qa_id"), []).append(ev)
scored = []
for rec in records:
evals = [e for e in evals_by_id.get(rec["qa_id"], [])
if e.get("judge") == judges[0]]
score = round(_trimmed_mean([e["score"] for e in evals]) * 100, 2) if evals else 0
scored.append({"qa_id": rec["qa_id"], "score": score})
total = len(scored)
mean = sum(s["score"] for s in scored) / total if total else 0
return {"mode": "rubric", "total": total,
"avg_score": round(mean, 2), "scored_records": scored}
def main():
parser = argparse.ArgumentParser(description="Rubric-score AudioSpan model outputs")
parser.add_argument("--input", required=True,
help="Prediction JSONL: {\"qa_id\": ..., \"answer\": \"...\"} per line")
parser.add_argument("--output", help="Aggregated scored JSONL (default: <input>_scored.jsonl)")
parser.add_argument("--evals", help="Per-eval JSONL (default: <input>_evals.jsonl)")
parser.add_argument("--data-root", type=Path, default=RELEASE_ROOT,
help="release root holding metadata/ (default: parent of evaluate/)")
parser.add_argument("--judges", help="Comma-separated judge models")
parser.add_argument("--rounds", type=int, default=DEFAULT_ROUNDS)
parser.add_argument("--judge-prompt", default=str(DEFAULT_JUDGE_PROMPT))
parser.add_argument("--api-base",
help="OpenAI-compatible base URL (default: inferred from first judge)")
parser.add_argument("--api-key",
help="API key (default: provider env var, see module docstring)")
args = parser.parse_args()
records = load_records(args.input, args.data_root.resolve())
judges = args.judges.split(",") if args.judges else DEFAULT_JUDGES
default_base, default_env = _resolve_default_api_base(judges[0])
if not args.api_base:
args.api_base = default_base
logger.info("Resolved --api-base from judge %r: %s", judges[0], args.api_base)
if not args.api_key:
args.api_key = os.environ.get(default_env) or os.environ.get("OPENAI_API_KEY")
if not args.api_key:
parser.error(f"--api-key or {default_env} required for judge {judges[0]!r}")
evals_path = args.evals or args.input.replace(".jsonl", "_evals.jsonl")
summary = run_rubric(
records, judges, args.rounds, evals_path,
judge_prompt_path=args.judge_prompt,
api_base=args.api_base, api_key=args.api_key,
)
out_path = args.output or args.input.replace(".jsonl", "_scored.jsonl")
with open(out_path, "w", encoding="utf-8") as fh:
for s in summary.get("scored_records", []):
fh.write(json.dumps(s, ensure_ascii=False) + "\n")
logger.info("Scored records: %s", out_path)
print(json.dumps({k: v for k, v in summary.items() if k != "scored_records"},
indent=2, ensure_ascii=False))
if __name__ == "__main__":
main()
|