Skip to content

v2.0.0 모델별 benchmark 결과

gannim edited this page Jul 30, 2025 · 1 revision

Total 

model judge type singlecall dialog calldecision overall score
claude-opus-4 LLM judge 0.9320 0.9800 0.9950 0.9690
Human judge 0.9160 0.9800 0.9950 0.9637
gemini-2.5-pro-preview-05-06 LLM judge 0.9300 0.9550 0.9769 0.9540
Human judge 0.9300 0.9450 0.9769 0.9506
Qwen3-32B(think) LLM judge 0.9280 0.9550 0.9670 0.9500
Human judge 0.9220 0.9550 0.9653 0.9474
gpt-4.1-2025-04-14 LLM judge 0.8980 0.9300 0.9092 0.9124
Human judge 0.9040 0.9200 0.8927 0.9056
gpt-4o-mini-2024-07-18 LLM judge 0.7820 0.8950 0.8564 0.8445
Human judge 0.7840 0.8950 0.8465 0.8418

SingleCall 

LLM judge rank model judge type exact (100) 4_random (100) 4_close (100) 8_random (100) 8_close (100) SUM (500) AVG (micro)
1 Qwen3-32B(think) LLM judge 0.9300 0.9300 0.9400 0.9400 0.9200 466 0.9320
Human judge 0.9100 0.9000 0.9300 0.9200 0.9200 458 0.9160
2 claude-opus-4 LLM judge 0.9300 0.9500 0.9500 0.9300 0.8900 465 0.9300
Human judge 0.9400 0.9400 0.9400 0.9300 0.9000 465 0.9300
3 gemini-2.5-pro-preview-05-06 LLM judge 0.9300 0.9200 0.9200 0.9400 0.9300 464 0.9280
Human judge 0.9400 0.9000 0.9100 0.9300 0.9300 461 0.9220
4 gpt-4.1-2025-04-14 LLM judge 0.9000 0.9100 0.9000 0.8900 0.8900 449 0.8980
Human judge 0.9200 0.9000 0.9100 0.8900 0.9000 452 0.9040
5 gpt-4o-mini-2024-07-18 LLM judge 0.8500 0.7700 0.7700 0.7700 0.7500 391 0.7820
Human judge 0.8500 0.7700 0.7700 0.7700 0.7600 392 0.7840

Dialog

LLM judge rank model judge type CALL (70) COMPLETION (71) SLOT (36) RELEVANCE (23) SUM (200) AVG (micro)
1 claude-opus-4 LLM judge 67 (0.9571) 71 (1.0000) 35 (0.9722) 23 (1.0000) 196 0.9800
Human judge 67 (0.9571) 71 (1.0000) 35 (0.9722) 23 (1.0000) 196 0.9800
2 gpt-4.1-2025-04-14 LLM judge 66 (0.9429) 68 (0.9577) 36 (1.0000) 21 (0.9130) 191 0.9550
Human judge 66 (0.9429) 67 (0.9437) 36 (1.0000) 20 (0.8696) 189 0.9450
3 gemini-2.5-pro-preview-05-06 LLM judge 65 (0.9286) 67 (0.9437) 36 (1.0000) 23 (1.0000) 191 0.9550
Human judge 65 (0.9286) 68 (0.9577) 35 (0.9722) 23 (1.0000) 191 0.9550
4 Qwen3-32B(think) LLM judge 68 (0.9714) 68 (0.9577) 32 (0.8889) 18 (0.7826) 186 0.9300
Human judge 68 (0.9714) 66 (0.9296) 33 (0.9167) 17 (0.7391) 184 0.9200
5 gpt-4o-mini-2024-07-18 LLM judge 63 (0.9000) 70 (0.9859) 34 (0.9444) 12 (0.5217) 179 0.8950
Human judge 63 (0.9000) 70 (0.9859) 33 (0.9167) 13 (0.5652) 179 0.8950

CallDecision

LLM judge rank model judge type CALL (100) REJECT (100) SLOT-all (100) SLOT-some (306) SUM (606) AVG (micro)
1 claude-opus-4 LLM judge 100 (1.0000) 100 (1.0000) 100 (1.0000) 303 (0.9902) 603 0.9950
Human judge 100 (1.0000) 100 (1.0000) 100 (1.0000) 303 (0.9902) 603 0.9950
2 gemini-2.5-pro-preview-05-06 LLM judge 96 (0.9600) 98 (0.9800) 96 (0.9600) 302 (0.9876) 592 0.9769
Human judge 97 (0.9700) 95 (0.9500) 98 (0.9800) 302 (0.9876) 592 0.9769
3 Qwen3-32B(think) LLM judge 94 (0.9400) 96 (0.9600) 98 (0.9800) 298 (0.9745) 586 0.9670
Human judge 99 (0.9900) 93 (0.9300) 97 (0.9700) 296 (0.9673) 585 0.9653
4 gpt-4.1-2025-04-14 LLM judge 93 (0.9300) 65 (0.6500) 100 (1.0000) 293 (0.9575) 551 0.9092
Human judge 94 (0.9400) 58 (0.5800) 100 (1.0000) 289 (0.9444) 541 0.8927
5 gpt-4o-mini-2024-07-18 LLM judge 87 (0.8700) 46 (0.4600) 100 (1.0000) 286 (0.9346) 519 0.8564
Human judge 87 (0.8700) 41 (0.4100) 100 (1.0000) 285 (0.9314) 513 0.8465

Clone this wiki locally