File size: 36,176 Bytes
9c60174 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 | import argparse
import json
import os
import re
import time
import numpy as np
from openai import OpenAI
from tqdm import tqdm
try:
from openai import AzureOpenAI
from azure.identity import (
AzureCliCredential,
ChainedTokenCredential,
ManagedIdentityCredential,
get_bearer_token_provider,
)
AZURE_OAUTH_SCOPE = os.environ.get("AZURE_OAUTH_SCOPE", "")
if AZURE_OAUTH_SCOPE:
credential = get_bearer_token_provider(
ChainedTokenCredential(
AzureCliCredential(),
ManagedIdentityCredential(),
),
AZURE_OAUTH_SCOPE,
)
else:
credential = None
except ImportError:
AzureOpenAI = None
credential = None
from model_zoo import model_zoo
# Azure OpenAI endpoint (set AZURE_OPENAI_ENDPOINT env var to your deployment URL).
endpoint = os.environ.get("AZURE_OPENAI_ENDPOINT", "")
# OpenAI-compatible LiteLLM proxy URL (set LITELLM_BASE_URL env var to your proxy).
TRITONAI_BASE_URL = os.environ.get("LITELLM_BASE_URL", "")
ATOMIC_PROMPT_VERSION = "atomic-v1"
LEGACY_PROMPT_VERSION = "binary-v0"
ATOM_SCORES = {
"correct": 1.0,
"partially_correct": 0.5,
"missing": 0.0,
"incorrect": 0.0,
}
def _retryable_status(e) -> "int | None":
status = getattr(e, "status_code", None) or getattr(e, "http_status", None)
if status in (429, 500, 503, 403):
return status
resp = getattr(e, "response", None)
if resp is not None and getattr(resp, "status_code", None) in (429, 500, 503):
return resp.status_code
msg = str(e).lower()
if "429" in msg or "rate limit" in msg:
return 429
if "500" in msg or "internal server error" in msg:
return 500
if "503" in msg or "api configuration unavailable" in msg:
return 503
return None
def parse_json_object(text):
text = (text or "").strip()
if text.startswith("```"):
text = re.sub(r"^```(?:json)?", "", text).strip()
text = re.sub(r"```$", "", text).strip()
try:
return json.loads(text)
except json.JSONDecodeError:
start = text.find("{")
end = text.rfind("}") + 1
if start >= 0 and end > start:
return json.loads(text[start:end])
raise
def sanitize_model_name(name):
return re.sub(r"[^A-Za-z0-9_.-]+", "_", name)
def read_json_or_jsonl(path):
try:
with open(path, "r", encoding="utf-8") as f:
return json.load(f)
except json.JSONDecodeError:
with open(path, "r", encoding="utf-8") as f:
return [json.loads(line) for line in f if line.strip()]
def read_existing_jsonl(path):
if not path or not os.path.exists(path):
return {}
rows = {}
with open(path, "r", encoding="utf-8") as f:
for line in f:
if not line.strip():
continue
obj = json.loads(line)
if "question_id" in obj:
rows[obj["question_id"]] = obj
return rows
def write_json(path, obj):
tmp_path = path + ".tmp"
with open(tmp_path, "w", encoding="utf-8") as f:
json.dump(obj, f, ensure_ascii=False, indent=2)
f.write("\n")
os.replace(tmp_path, path)
def append_jsonl(path, row):
with open(path, "a", encoding="utf-8") as f:
print(json.dumps(row, ensure_ascii=False), file=f, flush=True)
def default_result_file(hyp_file, metric_model, eval_mode):
if eval_mode == "legacy":
return f"{hyp_file}.eval-results-{metric_model}"
model_tag = sanitize_model_name(metric_model)
return f"{hyp_file}.eval-results-{model_tag}-{ATOMIC_PROMPT_VERSION}-{eval_mode}"
def default_rubric_file(ref_file):
return f"{ref_file}.{ATOMIC_PROMPT_VERSION}.rubric.json"
def load_rubric_file(path, ref_file):
if not path or not os.path.exists(path):
return {
"prompt_version": ATOMIC_PROMPT_VERSION,
"source_ref_file": ref_file,
"rubrics": {},
}
with open(path, "r", encoding="utf-8") as f:
data = json.load(f)
if "rubrics" in data:
data.setdefault("prompt_version", ATOMIC_PROMPT_VERSION)
data.setdefault("source_ref_file", ref_file)
return data
return {
"prompt_version": ATOMIC_PROMPT_VERSION,
"source_ref_file": ref_file,
"rubrics": data,
}
def question_type_guidance(task):
if task in ["Information Absence"]:
return (
"The correct answer is that the information is unavailable, absent, "
"not yet known, or not supported. A response that gives a concrete "
"answer instead of abstaining is incorrect."
)
if task in ["Aggregation", "single-session-user", "single-session-assistant", "multi-session"]:
return (
"Check every requested item, count, list member, and named fact. "
"Exact counts and required list coverage matter. Extra material "
"facts that change the answer should be flagged."
)
if task in ["Aggregation + Temporal"]:
return (
"Check both the aggregated facts and their time/order associations. "
"An answer can name the right item but still be wrong if the timing, "
"ordering, before/after relation, or year is wrong."
)
if task in ["Temporal Reasoning", "temporal-reasoning"]:
return (
"Check the specific time, date, year, sequence, duration, or temporal "
"relationship asked for. Accept +/-1 only for day/week/month duration "
"counts, not for years, event identity, or ordering."
)
if task in ["Knowledge Update", "knowledge-update"]:
return (
"Check the current or most recent state. Historical context is fine "
"only if the final/current state is clearly correct. Outdated states "
"presented as current are incorrect."
)
if task == "single-session-preference":
return (
"Check whether the response recalls and applies the stated preference. "
"Do not require unnecessary verbosity, but contradictions are incorrect."
)
return "Check whether the model response satisfies all required facts in the reference answer."
def get_anscheck_prompt(task, question, answer, response):
"""Legacy binary yes/no prompt kept for backward compatibility."""
if task in ["Information Absence"]:
template = """You are evaluating whether a model correctly identified that a question is unanswerable given the available personal chat history.
Question: {question}
Explanation of why it is unanswerable: {answer}
Model Response: {response}
Evaluation criteria:
- CORRECT if the model explicitly states that the information is not available, insufficient, or that the question cannot be answered from the provided context.
- INCORRECT if the model fabricates an answer or fails to acknowledge the unanswerable nature of the question.
- The model does not need to use the exact word "unanswerable" -- expressing uncertainty or lack of information is sufficient.
Briefly explain your reasoning (1-2 sentences), then on the last line write only: yes or no"""
elif task in ["Aggregation", "single-session-user", "single-session-assistant", "multi-session"]:
template = """You are evaluating whether a model correctly answered a question that requires aggregating specific facts from a user's personal chat history.
Question: {question}
Reference Answer: {answer}
Model Response: {response}
Evaluation criteria:
- CORRECT if the response identifies all key items, facts, or counts present in the reference answer, even if phrased differently or with added context.
- INCORRECT if the response:
- States a wrong count (e.g., says "5" when the answer is "3")
- Omits one or more key items/facts listed in the reference
- Lists mostly wrong items even if the count is right
- Partial answers that cover only a subset of required items are INCORRECT.
- Verbose responses are acceptable as long as all reference items are present within them.
- If the response contains correct items but also lists additional plausible-sounding but unverified items beyond the reference, this does NOT make it incorrect -- evaluate only whether the reference items are covered.
- Semantic equivalence counts as correct (e.g., "RSI" = "Repetitive Strain Injury").
Briefly explain your reasoning (2-3 sentences), then on the last line write only: yes or no"""
elif task in ["Aggregation + Temporal"]:
template = """You are evaluating whether a model correctly answered a question that requires both aggregating facts AND reasoning about their temporal order or time associations from a user's personal chat history.
Question: {question}
Reference Answer: {answer}
Model Response: {response}
Evaluation criteria:
- CORRECT if the response captures both:
(a) all key events/facts listed in the reference, and
(b) their correct temporal associations (ordering, time periods, or "before/after" relationships).
- INCORRECT if the response:
- Omits one or more key events or facts from the reference
- Gets the temporal ordering or time associations wrong
- Captures only the content without the temporal aspects, or vice versa
- Responses that describe the correct progression/sequence in different words are acceptable.
- Partial answers covering only some events or ignoring time aspects are INCORRECT.
- Minor wording differences or additional explanatory context are acceptable.
Briefly explain your reasoning (2-3 sentences), then on the last line write only: yes or no"""
elif task in ["Temporal Reasoning", "temporal-reasoning"]:
template = """You are evaluating whether a model correctly answered a question about temporal relationships in a user's personal chat history.
Question: {question}
Reference Answer: {answer}
Model Response: {response}
Evaluation criteria:
- CORRECT if the response correctly identifies the specific time, date, year, sequence, or temporal relationship asked about.
- INCORRECT if the response states a wrong year, wrong sequence, wrong temporal relationship, or misidentifies which event came first/last.
- Off-by-one tolerance: if the question asks for a count of days, weeks, or months, accept answers that differ by +/-1. This tolerance does NOT apply to years or to identifying specific events/artifacts.
- Responses that correctly identify the fact but with verbose context are acceptable.
- If the response hedges but still states the correct answer, it is correct.
Briefly explain your reasoning (2-3 sentences), then on the last line write only: yes or no"""
elif task in ["Knowledge Update", "knowledge-update"]:
template = """You are evaluating whether a model correctly answered a question about the most recent or current state of something that changed over time in a user's personal chat history.
Question: {question}
Reference Answer: {answer}
Model Response: {response}
Evaluation criteria:
- CORRECT if the response correctly identifies the most recent/current state as described in the reference answer.
- The response may include earlier historical states as context -- this is acceptable as long as the current/final state is correctly identified and clearly stated.
- INCORRECT if the response:
- States an outdated/superseded state as the current one
- Omits the current state entirely
- Correctly describes history but draws the wrong conclusion about what the current state is
- Semantic equivalence counts (e.g., "flexitarian" and "semi-vegetarian diet with occasional meat" are equivalent if contextually clear).
Briefly explain your reasoning (2-3 sentences), then on the last line write only: yes or no"""
elif task == "single-session-preference":
template = """You are evaluating whether a model correctly answered a personalized question based on a user's stated preferences from their chat history.
Question: {question}
Reference Rubric: {answer}
Model Response: {response}
Evaluation criteria:
- CORRECT if the response recalls and applies the user's personal preferences correctly, even if not covering every point in the rubric.
- INCORRECT if the response ignores, contradicts, or misremembers the user's preferences.
Briefly explain your reasoning (1-2 sentences), then on the last line write only: yes or no"""
else:
template = """You are evaluating whether a model's response correctly answers a question based on a user's personal chat history.
Question: {question}
Reference Answer: {answer}
Model Response: {response}
Is the response correct? It is correct if it contains all key information from the reference answer, even if phrased differently.
Briefly explain your reasoning (1-2 sentences), then on the last line write only: yes or no"""
return template.format(question=question, answer=answer, response=response)
def build_rubric_prompt(task, question, answer):
return f"""You are creating an atomic grading rubric for an open-ended QA benchmark.
Question type: {task}
Question:
{question}
Reference answer:
{answer}
Question-type guidance:
{question_type_guidance(task)}
Decompose the reference answer into the smallest independently checkable requirements needed to answer the question.
Rules:
- Each atom should be a single required answer unit: an entity, count, date/year, order relation, current-state conclusion, or abstention requirement.
- If an entity and its temporal relation are inseparable for correctness, keep them in the same atom.
- For list/count questions, include an atom for the exact count when the question asks "how many", and atoms for each required listed item when the item identities matter.
- For Information Absence, usually use one atom requiring the response to clearly state that the information is unavailable/insufficient/not discussed, and add a strict note that concrete fabricated answers are wrong.
- Do not include supporting evidence requirements or session IDs unless the question explicitly asks for them.
- Weights should normally be 1.0. Use a higher weight only when one atom is clearly the main answer and other atoms are minor.
Return JSON only with this schema:
{{
"required_atoms": [
{{
"id": "a1",
"requirement": "short, specific grading requirement",
"weight": 1.0
}}
],
"strict_notes": ["short note about exactness, ordering, abstention, or hallucination handling"]
}}"""
def build_atomic_eval_prompt(task, question, answer, response, rubric):
rubric_str = json.dumps(
{
"required_atoms": rubric["required_atoms"],
"strict_notes": rubric.get("strict_notes", []),
},
ensure_ascii=False,
indent=2,
)
return f"""You are an LLM-as-a-judge evaluating one model response against a reference answer.
Question type: {task}
Question:
{question}
Reference answer:
{answer}
Model response:
{response}
Atomic grading rubric:
{rubric_str}
Question-type guidance:
{question_type_guidance(task)}
Judge each atom independently.
Atom labels:
- correct: the response fully satisfies this atom, allowing semantic paraphrase.
- partially_correct: the response gets the main idea but is incomplete or slightly underspecified. Use this sparingly. Do not use it for wrong counts, wrong years, wrong named entities, wrong ordering, or a concrete answer to an Information Absence question.
- missing: the response does not address this atom.
- incorrect: the response contradicts this atom or gives the wrong count/entity/date/order/current state.
Also identify unsupported_or_contradictory material:
- severity "material": extra answer content that changes the final answer, adds extra items to an exact list/count, gives an outdated current state, fabricates a concrete answer for Information Absence, or contradicts any atom.
- severity "minor": harmless context or extra explanation that does not change the answer.
Return JSON only with this schema:
{{
"atom_judgments": [
{{
"id": "a1",
"label": "correct|partially_correct|missing|incorrect",
"rationale": "brief reason"
}}
],
"unsupported_or_contradictory": [
{{
"text": "extra or contradictory claim",
"severity": "minor|material",
"rationale": "brief reason"
}}
],
"absence_mismatch": false,
"overall_rationale": "one or two sentence summary"
}}"""
def llm_call(
deployment_name: str,
api_version: str,
_prompt: str,
debug: bool = False,
vllm: bool = False,
tritonai: bool = False,
nvidia: bool = False,
):
if nvidia:
client = OpenAI(
api_key=os.getenv("NV_API_KEY"),
base_url="https://inference-api.nvidia.com",
)
while True:
try:
return client.chat.completions.create(
model=deployment_name,
messages=[{"role": "user", "content": _prompt}],
)
except Exception as e:
st = _retryable_status(e)
if st in (429, 500, 503, 403):
print(f"[WARN] HTTP {st} from NVIDIA API; sleeping 60s then retrying...", flush=True)
time.sleep(60)
continue
print("One exception captured", repr(e), flush=True)
raise
if tritonai:
client = OpenAI(
api_key=os.getenv("TRITONAI_API_KEY"),
base_url=TRITONAI_BASE_URL,
)
while True:
try:
return client.chat.completions.create(
model=deployment_name,
messages=[{"role": "user", "content": _prompt}],
)
except Exception as e:
st = _retryable_status(e)
if st in (429, 500, 503, 403):
print(f"[WARN] HTTP {st} from LiteLLM proxy; sleeping 60s then retrying...", flush=True)
time.sleep(60)
continue
print("One exception captured", repr(e), flush=True)
raise
if deployment_name.startswith("claude-"):
import anthropic
client = anthropic.Anthropic(api_key=os.getenv("ANTHROPIC_API_KEY"))
while True:
try:
msg = client.messages.create(
model=deployment_name,
max_tokens=1024,
messages=[{"role": "user", "content": _prompt}],
)
class _Choice:
class _Msg:
def __init__(self, text):
self.content = text
def __init__(self, text):
self.message = self._Msg(text)
class _Completion:
def __init__(self, text):
self.choices = [_Choice(text)]
return _Completion(msg.content[0].text)
except Exception as e:
st = _retryable_status(e)
if st in (429, 500, 503, 403):
print(f"[WARN] HTTP {st} from Anthropic; sleeping 60s then retrying...", flush=True)
time.sleep(60)
continue
print("One exception captured", repr(e), flush=True)
raise
if vllm:
client = OpenAI(
base_url=os.getenv("VLLM_BASE_URL", "http://localhost:8000/v1"),
api_key=os.getenv("VLLM_API_KEY", "EMPTY"),
)
elif debug:
client = OpenAI(api_key=os.getenv("OPENAI_API_KEY"))
else:
client = AzureOpenAI(
azure_endpoint=endpoint,
azure_ad_token_provider=credential,
api_version=api_version,
)
kwargs = {
"model": deployment_name,
"messages": [{"role": "system", "content": _prompt}],
}
while True:
try:
return client.chat.completions.create(**kwargs)
except Exception as e:
st = _retryable_status(e)
if st in (429, 500, 503, 403):
print(f"[WARN] HTTP {st} from LLM; sleeping 120s then retrying...", flush=True)
time.sleep(120)
continue
print("One exception captured", repr(e), flush=True)
raise
def call_json_llm(prompt, deployment_name, api_version, args, max_retries=3):
last_error = None
for attempt in range(max_retries):
completion = llm_call(
deployment_name,
api_version,
prompt,
debug=args.debug,
vllm=args.vllm,
tritonai=args.tritonai,
nvidia=args.nvidia,
)
content = completion.choices[0].message.content.strip()
try:
return parse_json_object(content), content
except Exception as e:
last_error = e
if attempt < max_retries - 1:
print(f"[WARN] Failed to parse judge JSON; retrying ({attempt + 1}/{max_retries})", flush=True)
time.sleep(2)
raise ValueError(f"Failed to parse JSON response from judge: {last_error}")
def fallback_rubric(qid, task, question, answer):
return {
"question_id": qid,
"question_type": task,
"question": question,
"reference_answer": answer,
"required_atoms": [
{
"id": "a1",
"requirement": f"Response must correctly answer the question according to the reference answer: {answer}",
"weight": 1.0,
}
],
"strict_notes": ["Fallback single-atom rubric produced because the generated rubric was invalid."],
"prompt_version": ATOMIC_PROMPT_VERSION,
}
def normalize_rubric(qid, task, question, answer, parsed):
atoms = parsed.get("required_atoms", []) if isinstance(parsed, dict) else []
norm_atoms = []
for idx, atom in enumerate(atoms, start=1):
if not isinstance(atom, dict):
continue
requirement = str(atom.get("requirement", "")).strip()
if not requirement:
continue
atom_id = str(atom.get("id", f"a{idx}")).strip() or f"a{idx}"
try:
weight = float(atom.get("weight", 1.0))
except (TypeError, ValueError):
weight = 1.0
if weight <= 0:
weight = 1.0
norm_atoms.append({"id": atom_id, "requirement": requirement, "weight": weight})
if not norm_atoms:
return fallback_rubric(qid, task, question, answer)
strict_notes = parsed.get("strict_notes", [])
if not isinstance(strict_notes, list):
strict_notes = [str(strict_notes)]
return {
"question_id": qid,
"question_type": task,
"question": question,
"reference_answer": answer,
"required_atoms": norm_atoms,
"strict_notes": [str(x) for x in strict_notes],
"prompt_version": ATOMIC_PROMPT_VERSION,
}
def get_or_build_rubric(qdata, rubric_data, rubric_file, deployment_name, api_version, args):
qid = qdata["question_id"]
existing = rubric_data["rubrics"].get(qid)
if (
existing
and existing.get("prompt_version") == ATOMIC_PROMPT_VERSION
and existing.get("required_atoms")
and not args.force_rebuild_rubric
):
return existing
task = qdata["question_type"]
prompt = build_rubric_prompt(task, qdata["question"], qdata["answer"])
try:
parsed, raw = call_json_llm(prompt, deployment_name, api_version, args)
rubric = normalize_rubric(qid, task, qdata["question"], qdata["answer"], parsed)
rubric["rubric_raw_response"] = raw
except Exception as e:
print(f"[WARN] Falling back to single-atom rubric for {qid}: {e}", flush=True)
rubric = fallback_rubric(qid, task, qdata["question"], qdata["answer"])
rubric_data["rubrics"][qid] = rubric
write_json(rubric_file, rubric_data)
return rubric
def compute_atomic_scores(rubric, parsed):
atoms = rubric.get("required_atoms", [])
judgments_by_id = {}
raw_judgments = parsed.get("atom_judgments", []) if isinstance(parsed, dict) else []
if isinstance(raw_judgments, list):
for judgment in raw_judgments:
if not isinstance(judgment, dict):
continue
atom_id = str(judgment.get("id", "")).strip()
label = str(judgment.get("label", "")).strip()
if label not in ATOM_SCORES:
label = "incorrect"
judgments_by_id[atom_id] = {
"id": atom_id,
"label": label,
"score": ATOM_SCORES[label],
"rationale": str(judgment.get("rationale", "")),
}
norm_judgments = []
weighted_score = 0.0
total_weight = 0.0
for atom in atoms:
atom_id = atom["id"]
weight = float(atom.get("weight", 1.0))
judgment = judgments_by_id.get(
atom_id,
{"id": atom_id, "label": "missing", "score": 0.0, "rationale": "No judgment returned."},
)
judgment["requirement"] = atom["requirement"]
judgment["weight"] = weight
norm_judgments.append(judgment)
weighted_score += judgment["score"] * weight
total_weight += weight
extras = parsed.get("unsupported_or_contradictory", []) if isinstance(parsed, dict) else []
if not isinstance(extras, list):
extras = []
material_extras = [
x for x in extras
if isinstance(x, dict) and str(x.get("severity", "")).strip() == "material"
]
absence_mismatch = bool(parsed.get("absence_mismatch", False)) if isinstance(parsed, dict) else False
strict_label = (
bool(norm_judgments)
and all(j["label"] == "correct" for j in norm_judgments)
and not material_extras
and not absence_mismatch
)
partial_score = weighted_score / total_weight if total_weight > 0 else 0.0
if absence_mismatch:
partial_score = 0.0
elif material_extras and partial_score > 0.8:
partial_score = 0.8
return {
"strict_label": strict_label,
"partial_score": round(partial_score, 4),
"atom_judgments": norm_judgments,
"unsupported_or_contradictory": extras,
"absence_mismatch": absence_mismatch,
"overall_rationale": str(parsed.get("overall_rationale", "")) if isinstance(parsed, dict) else "",
}
def judge_atomic(qdata, hypothesis, rubric, deployment_name, api_version, args):
prompt = build_atomic_eval_prompt(
qdata["question_type"],
qdata["question"],
qdata["answer"],
hypothesis,
rubric,
)
parsed, raw = call_json_llm(prompt, deployment_name, api_version, args)
scores = compute_atomic_scores(rubric, parsed)
return {
"model": args.eval_model_name,
"prompt_version": ATOMIC_PROMPT_VERSION,
"eval_mode": args.eval_mode,
"strict_label": scores["strict_label"],
"partial_score": scores["partial_score"],
"required_atoms": rubric["required_atoms"],
"atom_judgments": scores["atom_judgments"],
"unsupported_or_contradictory": scores["unsupported_or_contradictory"],
"absence_mismatch": scores["absence_mismatch"],
"overall_rationale": scores["overall_rationale"],
"raw_response": raw,
}
def should_skip_existing(existing_row, eval_mode):
if eval_mode == "legacy":
return "autoeval_label" in existing_row
atomic = existing_row.get("autoeval_atomic")
return bool(atomic and atomic.get("prompt_version") == ATOMIC_PROMPT_VERSION)
def safe_mean(values):
if not values:
return float("nan")
return float(np.mean(values))
def print_legacy_summary(logs, qtype2acc):
labels = [1 if x["autoeval_label"]["label"] else 0 for x in logs if "autoeval_label" in x]
print("Accuracy:", round(safe_mean(labels), 4))
for k, v in sorted(qtype2acc.items()):
print("\t{}: {} ({})".format(k, round(safe_mean(v), 4), len(v)))
def print_atomic_summary(logs, qtype2strict, qtype2partial, eval_mode):
strict_values = [
1 if x["autoeval_atomic"]["strict_label"] else 0
for x in logs
if "autoeval_atomic" in x
]
partial_values = [
float(x["autoeval_atomic"]["partial_score"])
for x in logs
if "autoeval_atomic" in x
]
if eval_mode in ("strict", "both"):
print("Strict Accuracy:", round(safe_mean(strict_values), 4))
for k, v in sorted(qtype2strict.items()):
print("\t{}: {} ({})".format(k, round(safe_mean(v), 4), len(v)))
if eval_mode in ("partial", "both"):
print("Partial Score:", round(safe_mean(partial_values), 4))
for k, v in sorted(qtype2partial.items()):
print("\t{}: {} ({})".format(k, round(safe_mean(v), 4), len(v)))
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--hyp_file", type=str, default=None)
parser.add_argument("--ref_file", type=str, required=True)
parser.add_argument("--eval_model_name", type=str, required=True)
parser.add_argument(
"--eval_mode",
type=str,
default="both",
choices=["legacy", "strict", "partial", "both"],
help="legacy uses the old yes/no judge; strict/partial/both use atomic JSON judging.",
)
parser.add_argument("--rubric_file", type=str, default=None)
parser.add_argument("--build_rubric_only", action="store_true", default=False)
parser.add_argument("--force_rebuild_rubric", action="store_true", default=False)
parser.add_argument("--result_file", type=str, default=None)
parser.add_argument("--debug", action="store_true", default=False)
parser.add_argument("--vllm", action="store_true", default=False)
parser.add_argument("--tritonai", action="store_true", default=False,
help="Use OpenAI-compatible LiteLLM proxy (set TRITONAI_API_KEY env var)")
parser.add_argument("--nvidia", action="store_true", default=False,
help="Use NVIDIA inference API (set NV_API_KEY env var)")
parser.add_argument("--verbose", action=argparse.BooleanOptionalAction, default=True)
args = parser.parse_args()
if not args.build_rubric_only and not args.hyp_file:
parser.error("--hyp_file is required unless --build_rubric_only is set")
metric_model = args.eval_model_name
deployment_name, api_version = model_zoo[metric_model]
references = read_json_or_jsonl(args.ref_file)
qid2qdata = {entry["question_id"]: entry for entry in references}
qid2qtype = {entry["question_id"]: entry["question_type"] for entry in references}
qtypes = set(qid2qtype.values())
rubric_data = None
rubric_file = args.rubric_file or default_rubric_file(args.ref_file)
if args.eval_mode != "legacy" or args.build_rubric_only:
rubric_data = load_rubric_file(rubric_file, args.ref_file)
if args.build_rubric_only:
for entry in tqdm(references, desc="building rubrics"):
get_or_build_rubric(entry, rubric_data, rubric_file, deployment_name, api_version, args)
print(f"Saved rubric file to {rubric_file}")
return
result_file = args.result_file or default_result_file(args.hyp_file, metric_model, args.eval_mode)
existing = read_existing_jsonl(result_file)
hypotheses = read_json_or_jsonl(args.hyp_file)
qtype2acc = {t: [] for t in qtypes}
qtype2strict = {t: [] for t in qtypes}
qtype2partial = {t: [] for t in qtypes}
logs = []
for entry in tqdm(hypotheses):
qid = entry.get("question_id")
if qid not in qid2qtype:
if qid is not None:
print(f"Warning: skipping {qid} as it is not in reference data.")
continue
if qid in existing and should_skip_existing(existing[qid], args.eval_mode):
existing_row = existing[qid]
logs.append(existing_row)
qtype = qid2qtype[qid]
if args.eval_mode == "legacy":
label = existing_row["autoeval_label"]["label"]
qtype2acc[qtype].append(1 if label else 0)
else:
atomic = existing_row["autoeval_atomic"]
qtype2strict[qtype].append(1 if atomic["strict_label"] else 0)
qtype2partial[qtype].append(float(atomic["partial_score"]))
continue
qdata = qid2qdata[qid]
qtype = qdata["question_type"]
hyp = entry["hypothesis"]
if args.eval_mode == "legacy":
prompt = get_anscheck_prompt(qtype, qdata["question"], qdata["answer"], hyp)
completion = llm_call(
deployment_name,
api_version,
prompt,
debug=args.debug,
vllm=args.vllm,
tritonai=args.tritonai,
nvidia=args.nvidia,
)
eval_response = completion.choices[0].message.content.strip()
last_line = next((l.strip().lower() for l in reversed(eval_response.splitlines()) if l.strip()), "")
label = last_line == "yes" or last_line.startswith("yes")
row = dict(entry)
row["autoeval_label"] = {
"model": metric_model,
"prompt_version": LEGACY_PROMPT_VERSION,
"label": label,
"raw_response": eval_response,
}
logs.append(row)
qtype2acc[qtype].append(1 if label else 0)
if args.verbose:
print(json.dumps({
"question": qdata["question"],
"answer": qdata["answer"],
"hypothesis": hyp,
"autoeval_label": label,
}, indent=4), flush=True)
append_jsonl(result_file, row)
continue
rubric = get_or_build_rubric(qdata, rubric_data, rubric_file, deployment_name, api_version, args)
try:
atomic_eval = judge_atomic(qdata, hyp, rubric, deployment_name, api_version, args)
except ValueError as _judge_err:
print(f"[WARN] judge_atomic failed for {qdata['question_id']}, writing zero score: {_judge_err}", flush=True)
atoms = rubric.get("required_atoms", [])
atomic_eval = {
"model": deployment_name,
"prompt_version": ATOMIC_PROMPT_VERSION,
"eval_mode": args.eval_mode,
"strict_label": False,
"partial_score": 0.0,
"required_atoms": atoms,
"atom_judgments": [{"id": a["id"], "label": "error", "score": 0.0, "rationale": "judge parse error", "requirement": a.get("requirement", ""), "weight": a.get("weight", 1.0)} for a in atoms],
"unsupported_or_contradictory": [],
"absence_mismatch": False,
"overall_rationale": f"Skipped: judge JSON parse error ({_judge_err})",
}
row = dict(entry)
row["autoeval_atomic"] = atomic_eval
logs.append(row)
qtype2strict[qtype].append(1 if atomic_eval["strict_label"] else 0)
qtype2partial[qtype].append(float(atomic_eval["partial_score"]))
if args.verbose:
print(json.dumps({
"question": qdata["question"],
"answer": qdata["answer"],
"hypothesis": hyp,
"strict_label": atomic_eval["strict_label"],
"partial_score": atomic_eval["partial_score"],
"atom_judgments": atomic_eval["atom_judgments"],
}, indent=4), flush=True)
append_jsonl(result_file, row)
if args.eval_mode == "legacy":
print_legacy_summary(logs, qtype2acc)
else:
print_atomic_summary(logs, qtype2strict, qtype2partial, args.eval_mode)
print(f"Rubric file: {rubric_file}")
print("Saved to", result_file)
if __name__ == "__main__":
main()
|