File size: 24,554 Bytes
5db30ac
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import asyncio
import json
import sys
import unittest
from pathlib import Path

ROOT = Path(__file__).resolve().parents[1]
SRC = ROOT / "src"
if str(SRC) not in sys.path:
    sys.path.insert(0, str(SRC))

from elbench.judges.judge_highlevel import HighLevelJudge
from elbench.judges.judge_ifeval import IFEvalJudge
from elbench.judges.judge_math import AIMEJudge, Math500Judge
from elbench.judges.judge_mmlu import MMLUProJudge
from elbench.judges.judge_teaching_harm import TeachingHarmJudge
from elbench.judges.llm_judge import LLMJudge
from elbench.judges.router import JudgeRouter
from elbench.judges.ifeval_rule.instructions_registry import INSTRUCTION_DICT
from elbench.config import load_project_config
from elbench.registry import FileRegistry
from elbench.schemas.config import JudgeTaskConfig
from elbench.schemas.evaluation import ModelResponse, Sample
from elbench.utils import extract_json_object


class JudgeSmokeTest(unittest.TestCase):
    def test_llm_judge_returns_structured_result(self) -> None:
        config = load_project_config(Path("configs"))
        judge = LLMJudge(
            config,
            JudgeTaskConfig(mode="llm", template="highlevel_edu", judge_model_id="mock.judge"),
        )
        sample = Sample(
            sample_id="l1",
            source_file="高阶育人-edu.jsonl",
            source_path="高阶育人-edu.jsonl",
            module="高阶育人",
            subset="edu",
            task="highlevel_edu",
            prompt='Please provide "Score", "Scoring Details", and "Personalized Feedback" in JSON format.',
            metadata={"_scene": "AG"},
        )
        response = ModelResponse(text='{"Score": "4/5", "Scoring Details": "ok", "Personalized Feedback": "ok"}')
        result = asyncio.run(judge.judge(sample, response))
        self.assertEqual(result.judge_result, "pass")
        self.assertEqual(result.score, 1.0)

    def test_router_uses_llm_judge_for_subjective_tasks(self) -> None:
        config = load_project_config(Path("configs"))
        router = JudgeRouter(config)
        sample = Sample(
            sample_id="r1",
            source_file="安全拒答.jsonl",
            source_path="安全拒答.jsonl",
            module="安全可信",
            subset="通用-应拒答",
            task="safety_refusal",
            prompt="dummy",
        )
        judge = router.get_judge(sample)
        self.assertIsInstance(judge, LLMJudge)

    def test_router_uses_rule_judge_for_highlevel_edu(self) -> None:
        config = load_project_config(Path("configs"))
        router = JudgeRouter(config)
        sample = Sample(
            sample_id="edu1",
            source_file="高阶育人-edu.jsonl",
            source_path="高阶育人-edu.jsonl",
            module="高阶育人",
            subset="edu",
            task="highlevel_edu",
            prompt="dummy",
        )

        judge = router.get_judge(sample)

        self.assertIsInstance(judge, HighLevelJudge)

    def test_satas_exact_match(self) -> None:
        judge = TeachingHarmJudge()
        sample = Sample(
            sample_id="s1",
            source_file="SATAs.xlsx",
            source_path="SATAs.xlsx",
            module="安全可信",
            subset="教育-教学安全",
            task="teaching_harm",
            prompt="dummy",
            reference={"Answer": "A,B,E"},
        )
        response = ModelResponse(text="A, B, E")
        result = asyncio.run(judge.judge(sample, response))
        self.assertEqual(result.judge_result, "pass")
        self.assertEqual(result.score, 1.0)
        self.assertFalse(result.judge_metadata["partial_credit"])

    def test_highlevel_omni_exact_match(self) -> None:
        judge = HighLevelJudge()
        sample = Sample(
            sample_id="o1",
            source_file="高阶育人-omni.jsonl",
            source_path="高阶育人-omni.jsonl",
            module="高阶育人",
            subset="omni",
            task="highlevel_omni",
            prompt="dummy",
            reference={"answer": "C", "reasonableness": 1},
        )
        response = ModelResponse(text="答案:C")
        result = asyncio.run(judge.judge(sample, response))
        self.assertEqual(result.judge_result, "pass")
        self.assertEqual(result.score, 1.0)
        self.assertEqual(result.judge_metadata["predicted_answer"], "C")

    def test_satas_missing_option_receives_partial_credit(self) -> None:
        judge = TeachingHarmJudge()
        sample = Sample(
            sample_id="s2",
            source_file="SATAs.xlsx",
            source_path="SATAs.xlsx",
            module="安全可信",
            subset="教育-教学安全",
            task="teaching_harm",
            prompt="dummy",
            reference={"Answer": "A,B,E"},
        )
        response = ModelResponse(text="A, E")

        result = asyncio.run(judge.judge(sample, response))

        self.assertEqual(result.judge_result, "fail")
        self.assertEqual(result.score, 0.5)
        self.assertTrue(result.judge_metadata["partial_credit"])

    def test_satas_wrong_option_receives_zero_credit(self) -> None:
        judge = TeachingHarmJudge()
        sample = Sample(
            sample_id="s3",
            source_file="SATAs.xlsx",
            source_path="SATAs.xlsx",
            module="安全可信",
            subset="教育-教学安全",
            task="teaching_harm",
            prompt="dummy",
            reference={"Answer": "A,B,E"},
        )
        response = ModelResponse(text="A, C")

        result = asyncio.run(judge.judge(sample, response))

        self.assertEqual(result.judge_result, "fail")
        self.assertEqual(result.score, 0.0)
        self.assertFalse(result.judge_metadata["partial_credit"])

    def test_highlevel_omni_multiple_choice_exact_match(self) -> None:
        judge = HighLevelJudge()
        sample = Sample(
            sample_id="o_multi",
            source_file="高阶育人-omni.jsonl",
            source_path="高阶育人-omni.jsonl",
            module="高阶育人",
            subset="omni",
            task="highlevel_omni",
            prompt="dummy",
            reference={"answer": "AC", "reasonableness": 1},
        )
        response = ModelResponse(text="答案:AC")
        result = asyncio.run(judge.judge(sample, response))
        self.assertEqual(result.judge_result, "pass")
        self.assertEqual(result.score, 1.0)
        self.assertEqual(result.judge_metadata["predicted_answer"], "AC")
        self.assertEqual(result.judge_metadata["expected_answer"], "AC")

    def test_llm_judge_respects_model_max_tokens_cap(self) -> None:
        config = load_project_config(Path("configs"))
        judge = LLMJudge(
            config,
            JudgeTaskConfig(mode="llm", template="safety_refusal", judge_model_id="qwen-test"),
        )

        class DummyClient:
            def __init__(self) -> None:
                self.request = None

            async def generate(self, sample, request):
                self.request = request
                return ModelResponse(text='{"judge_result": "pass", "score": 1, "judge_reason": "ok"}')

            async def aclose(self) -> None:
                return None

        dummy = DummyClient()
        judge._client = dummy
        sample = Sample(
            sample_id="l2",
            source_file="安全拒答.jsonl",
            source_path="安全拒答.jsonl",
            module="安全可信",
            subset="通用-应拒答",
            task="safety_refusal",
            prompt="dummy",
        )
        response = ModelResponse(text="refuse")

        asyncio.run(judge.judge(sample, response))
        self.assertIsNotNone(dummy.request)
        self.assertEqual(dummy.request.max_tokens, 1024)
        self.assertLessEqual(dummy.request.max_tokens, config.models["qwen-test"].max_tokens)

    def test_highlevel_edu_ag_json_scoring(self) -> None:
        judge = HighLevelJudge()
        sample = Sample(
            sample_id="ag1",
            source_file="高阶育人-edu.jsonl",
            source_path="高阶育人-edu.jsonl",
            module="高阶育人",
            subset="edu",
            task="highlevel_edu",
            prompt='Please provide "Score", "Scoring Details", and "Personalized Feedback" in JSON format.',
            reference={
                "Score": "4/5",
                "Scoring Details": "The answer is mostly correct but could be more specific.",
                "Personalized Feedback": "Good work. Add one concrete example next time.",
            },
            metadata={"_scene": "AG"},
        )
        response = ModelResponse(
            text='{"Score": "4/5", "Scoring Details": "The answer is mostly correct but could be more specific.", "Personalized Feedback": "Good work. Add one concrete example next time."}'
        )
        result = asyncio.run(judge.judge(sample, response))
        self.assertEqual(result.judge_result, "pass")
        self.assertGreaterEqual(result.score or 0, 0.99)

    def test_mmlu_pro_explicit_answer_match(self) -> None:
        judge = MMLUProJudge()
        sample = self._mmlu_sample("D")
        response = ModelResponse(text="Reasoning here.\nANSWER: D")
        result = asyncio.run(judge.judge(sample, response))
        self.assertEqual(result.judge_result, "pass")
        self.assertEqual(result.score, 1.0)

    def test_extract_json_object_handles_nested_fenced_json(self) -> None:
        parsed = extract_json_object(
            '```json\n{"Answer": {"CorrectOption": "C", "Explanation": "ok"}}\n```'
        )

        self.assertEqual(parsed, {"Answer": {"CorrectOption": "C", "Explanation": "ok"}})

    def test_mmlu_pro_uses_last_explicit_answer(self) -> None:
        judge = MMLUProJudge()
        sample = self._mmlu_sample("B")
        response = ModelResponse(text="A seems tempting at first.\nANSWER: B")
        result = asyncio.run(judge.judge(sample, response))
        self.assertEqual(result.judge_result, "pass")
        self.assertEqual(result.score, 1.0)
        self.assertEqual(result.judge_metadata["predicted_answer"], "B")

    def test_mmlu_pro_accepts_answer_letter_with_option_text(self) -> None:
        judge = MMLUProJudge()
        sample = self._mmlu_sample("B")
        response = ModelResponse(text="Reasoning here.\nANSWER: B) negative")
        result = asyncio.run(judge.judge(sample, response))
        self.assertEqual(result.judge_result, "pass")
        self.assertEqual(result.score, 1.0)
        self.assertEqual(result.judge_metadata["predicted_answer"], "B")

    def test_mmlu_pro_wrong_answer_fails(self) -> None:
        judge = MMLUProJudge()
        sample = self._mmlu_sample("D")
        response = ModelResponse(text="ANSWER: E")
        result = asyncio.run(judge.judge(sample, response))
        self.assertEqual(result.judge_result, "fail")
        self.assertEqual(result.score, 0.0)

    def test_mmlu_pro_supports_j_option(self) -> None:
        judge = MMLUProJudge()
        sample = self._mmlu_sample("J")
        response = ModelResponse(text="ANSWER: J")
        result = asyncio.run(judge.judge(sample, response))
        self.assertEqual(result.judge_result, "pass")
        self.assertEqual(result.score, 1.0)

    def test_ceval_chinese_answer_prefix(self) -> None:
        judge = MMLUProJudge()
        sample = self._mmlu_sample("C", task="ceval", choices="ABCD")
        response = ModelResponse(text="逐步分析后可知,最后答案如下。\n答案:C")
        result = asyncio.run(judge.judge(sample, response))
        self.assertEqual(result.judge_result, "pass")
        self.assertEqual(result.score, 1.0)
        self.assertEqual(result.judge_metadata["predicted_answer"], "C")

    def test_ceval_accepts_answer_letter_with_option_text(self) -> None:
        judge = MMLUProJudge()
        sample = self._mmlu_sample("C", task="ceval", choices="ABCD")
        response = ModelResponse(text="逐步分析后可知,最后答案如下。\n答案:C)正确选项")
        result = asyncio.run(judge.judge(sample, response))
        self.assertEqual(result.judge_result, "pass")
        self.assertEqual(result.score, 1.0)
        self.assertEqual(result.judge_metadata["predicted_answer"], "C")

    def test_mmlu_pro_requires_explicit_final_answer_format(self) -> None:
        judge = MMLUProJudge()
        sample = self._mmlu_sample("D", choices="ABCD")
        response = ModelResponse(text="Options: A B C D")
        result = asyncio.run(judge.judge(sample, response))
        self.assertEqual(result.judge_result, "fail")
        self.assertIsNone(result.judge_metadata["predicted_answer"])

    def test_mmlu_pro_multiple_letters_without_final_marker_fails(self) -> None:
        judge = MMLUProJudge()
        sample = self._mmlu_sample("D", choices="ABCD")
        response = ModelResponse(text="I considered A, B, C, and D. D seems possible.")
        result = asyncio.run(judge.judge(sample, response))
        self.assertEqual(result.judge_result, "fail")
        self.assertIsNone(result.judge_metadata["predicted_answer"])

    def test_highlevel_omni_requires_explicit_final_answer_format(self) -> None:
        judge = HighLevelJudge()
        sample = Sample(
            sample_id="o2",
            source_file="高阶育人-omni.jsonl",
            source_path="高阶育人-omni.jsonl",
            module="高阶育人",
            subset="omni",
            task="highlevel_omni",
            prompt="dummy",
            reference={"answer": "C", "reasonableness": 1},
        )
        response = ModelResponse(text="A is tempting, but C is ultimately better.")
        result = asyncio.run(judge.judge(sample, response))
        self.assertEqual(result.judge_result, "fail")
        self.assertIsNone(result.judge_metadata["predicted_answer"])

    def test_highlevel_omni_does_not_accept_earlier_correct_letter(self) -> None:
        judge = HighLevelJudge()
        sample = Sample(
            sample_id="o3",
            source_file="高阶育人-omni.jsonl",
            source_path="高阶育人-omni.jsonl",
            module="高阶育人",
            subset="omni",
            task="highlevel_omni",
            prompt="dummy",
            reference={"answer": "C", "reasonableness": 1},
        )
        response = ModelResponse(text="C is one possible answer, but after reconsidering I choose A.")
        result = asyncio.run(judge.judge(sample, response))
        self.assertEqual(result.judge_result, "fail")
        self.assertIsNone(result.judge_metadata["predicted_answer"])

    def test_ifeval_no_comma_passes_and_fails(self) -> None:
        judge = IFEvalJudge()
        sample = self._ifeval_sample(["punctuation:no_comma"], [{}])

        pass_result = asyncio.run(judge.judge(sample, ModelResponse(text="No comma appears here")))
        fail_result = asyncio.run(judge.judge(sample, ModelResponse(text="A comma, appears here")))

        self.assertEqual(pass_result.judge_result, "pass")
        self.assertEqual(pass_result.score, 1.0)
        self.assertEqual(fail_result.judge_result, "fail")
        self.assertEqual(fail_result.score, 0.0)

    def test_ifeval_highlighted_sections_counts_single_and_double_markdown(self) -> None:
        judge = IFEvalJudge()
        sample = self._ifeval_sample(
            ["detectable_format:number_highlighted_sections"],
            [{"num_highlights": 3}],
        )
        response = ModelResponse(text="*one* **two** **   ** *three*")
        result = asyncio.run(judge.judge(sample, response))

        self.assertEqual(result.judge_result, "pass")
        self.assertEqual(result.score, 1.0)

    def test_ifeval_number_words_relations(self) -> None:
        judge = IFEvalJudge()
        at_least = self._ifeval_sample(["length_constraints:number_words"], [{"relation": "at least", "num_words": 3}])
        less_than = self._ifeval_sample(["length_constraints:number_words"], [{"relation": "less than", "num_words": 3}])

        self.assertEqual(asyncio.run(judge.judge(at_least, ModelResponse(text="one two three"))).score, 1.0)
        self.assertEqual(asyncio.run(judge.judge(less_than, ModelResponse(text="one two"))).score, 1.0)
        self.assertEqual(asyncio.run(judge.judge(less_than, ModelResponse(text="one two three"))).score, 0.0)

    def test_ifeval_number_placeholders(self) -> None:
        judge = IFEvalJudge()
        sample = self._ifeval_sample(["detectable_content:number_placeholders"], [{"num_placeholders": 2}])
        result = asyncio.run(judge.judge(sample, ModelResponse(text="Hello [name] at [address].")))

        self.assertEqual(result.judge_result, "pass")
        self.assertEqual(result.score, 1.0)

    def test_ifeval_multi_instruction_averages_instruction_level(self) -> None:
        judge = IFEvalJudge()
        sample = self._ifeval_sample(
            [
                "punctuation:no_comma",
                "detectable_content:number_placeholders",
                "length_constraints:number_words",
            ],
            [{}, {"num_placeholders": 1}, {"relation": "at least", "num_words": 100}],
        )
        result = asyncio.run(judge.judge(sample, ModelResponse(text="No comma here [slot]")))

        self.assertEqual(result.judge_result, "fail")
        self.assertEqual(result.judge_metadata["prompt_level_strict"], 0.0)
        self.assertAlmostEqual(result.judge_metadata["inst_level_strict"], 2 / 3)

    def test_ifeval_loose_mode_tries_trimmed_lines(self) -> None:
        judge = IFEvalJudge()
        sample = self._ifeval_sample(["punctuation:no_comma"], [{}])
        result = asyncio.run(judge.judge(sample, ModelResponse(text="bad, first line\nclean final line")))

        self.assertEqual(result.judge_metadata["prompt_level_strict"], 0.0)
        self.assertEqual(result.judge_metadata["prompt_level_loose"], 1.0)

    def test_ifeval_dataset_instructions_are_all_registered(self) -> None:
        config = load_project_config(Path("configs"))
        registry = FileRegistry(config)
        resolved = registry.resolve(source_files={"ifeval_sampled.jsonl"})
        self.assertEqual(len(resolved), 1)

        instruction_ids: set[str] = set()
        with resolved[0].path.open("r", encoding="utf-8") as handle:
            for line in handle:
                if not line.strip():
                    continue
                record = json.loads(line)
                metadata = record.get("metadata") or {}
                raw_instruction_ids = metadata.get("instruction_id_list") or []
                instruction_ids.update(str(item) for item in raw_instruction_ids)

        self.assertTrue(instruction_ids)
        self.assertTrue(instruction_ids.issubset(set(INSTRUCTION_DICT)))

    def test_math_500_extracts_boxed_answer(self) -> None:
        judge = Math500Judge()
        sample = self._math_sample("42")
        result = asyncio.run(judge.judge(sample, ModelResponse(text="Reasoning here. Therefore, \\boxed{42}.")))

        self.assertEqual(result.judge_result, "pass")
        self.assertEqual(result.score, 1.0)
        self.assertEqual(result.judge_metadata["predicted_answer"], "42")

    def test_math_500_extracts_answer_marker_and_last_number(self) -> None:
        judge = Math500Judge()

        marker_result = asyncio.run(judge.judge(self._math_sample("9"), ModelResponse(text="The answer is 9.")))
        fallback_result = asyncio.run(judge.judge(self._math_sample("-50"), ModelResponse(text="After simplification we get -50.")))

        self.assertEqual(marker_result.score, 1.0)
        self.assertEqual(fallback_result.score, 1.0)

    def test_math_500_normalizes_latex_fraction_and_text(self) -> None:
        judge = Math500Judge()

        fraction_result = asyncio.run(judge.judge(self._math_sample("\\frac{14}{3}"), ModelResponse(text="\\boxed{14/3}")))
        text_result = asyncio.run(judge.judge(self._math_sample("\\text{Evelyn}"), ModelResponse(text="final answer is Evelyn")))

        self.assertEqual(fraction_result.score, 1.0)
        self.assertEqual(text_result.score, 1.0)

    def test_math_500_tuple_latex_matches_reference(self) -> None:
        judge = Math500Judge()
        sample = self._math_sample("\\left( 3, \\frac{\\pi}{2} \\right)")
        response = ModelResponse(text="The polar form is \\boxed{\\left(3,\\frac{\\pi}{2}\\right)}.")
        result = asyncio.run(judge.judge(sample, response))

        self.assertEqual(result.judge_result, "pass")
        self.assertEqual(result.score, 1.0)

    def test_math_500_wrong_answer_fails(self) -> None:
        judge = Math500Judge()
        result = asyncio.run(judge.judge(self._math_sample("42"), ModelResponse(text="\\boxed{43}")))

        self.assertEqual(result.judge_result, "fail")
        self.assertEqual(result.score, 0.0)

    def test_aime_extracts_boxed_reference_and_prediction(self) -> None:
        judge = AIMEJudge()
        result = asyncio.run(judge.judge(self._aime_sample("\\boxed{204}"), ModelResponse(text="After solving, \\boxed{204}")))

        self.assertEqual(result.judge_result, "pass")
        self.assertEqual(result.score, 1.0)
        self.assertEqual(result.judge_metadata["expected_answer"], "204")
        self.assertEqual(result.judge_metadata["predicted_answer"], "204")

    def test_aime_uses_last_number_fallback_for_prediction(self) -> None:
        judge = AIMEJudge()
        result = asyncio.run(judge.judge(self._aime_sample("\\boxed{113}"), ModelResponse(text="The requested value is 113.")))

        self.assertEqual(result.judge_result, "pass")
        self.assertEqual(result.score, 1.0)

    def test_aime_wrong_answer_fails(self) -> None:
        judge = AIMEJudge()
        result = asyncio.run(judge.judge(self._aime_sample("\\boxed{371}"), ModelResponse(text="\\boxed{370}")))

        self.assertEqual(result.judge_result, "fail")
        self.assertEqual(result.score, 0.0)

    def test_aime25_and_aime26_use_same_rule_judge(self) -> None:
        config = load_project_config(Path("configs"))
        router = JudgeRouter(config)

        aime25 = self._aime_sample("70", task="aime25", source_file="aime25.jsonl")
        aime26 = self._aime_sample("277", task="aime26", source_file="aime26.jsonl")

        self.assertIsInstance(router.get_judge(aime25), AIMEJudge)
        self.assertIsInstance(router.get_judge(aime26), AIMEJudge)

    def _mmlu_sample(self, target: str, task: str = "mmlu_pro", choices: str = "ABCDEFGHIJ") -> Sample:
        return Sample(
            sample_id="mmlu1",
            source_file=f"{task}_sampled.jsonl",
            source_path=f"{task}_sampled.jsonl",
            module="通用模型",
            subset=task,
            task=task,
            dimension="law",
            prompt="dummy",
            reference={"target": target},
            metadata={"choices": [f"choice {letter}" for letter in choices]},
        )

    def _ifeval_sample(self, instruction_ids: list[str], kwargs: list[dict]) -> Sample:
        return Sample(
            sample_id="ifeval1",
            source_file="ifeval_sampled.jsonl",
            source_path="ifeval_sampled.jsonl",
            module="通用模型",
            subset="ifeval",
            task="ifeval",
            dimension="default",
            prompt="dummy",
            reference={"target": ""},
            metadata={
                "instruction_id_list": instruction_ids,
                "kwargs": kwargs,
                "key": 1,
            },
        )

    def _math_sample(self, target: str) -> Sample:
        return Sample(
            sample_id="math1",
            source_file="math_500_sampled.jsonl",
            source_path="math_500_sampled.jsonl",
            module="通用模型",
            subset="math_500",
            task="math_500",
            dimension="Level 2",
            prompt="dummy",
            reference={"target": target},
            metadata={"question_id": "test/example.json", "subset_key": "Level 2"},
        )

    def _aime_sample(self, target: str, task: str = "aime24", source_file: str = "aime24_sampled.jsonl") -> Sample:
        return Sample(
            sample_id="aime1",
            source_file=source_file,
            source_path=source_file,
            module="通用模型",
            subset=task,
            task=task,
            dimension="default",
            prompt="dummy",
            reference={"target": target},
            metadata={},
        )


if __name__ == "__main__":
    unittest.main()