yjoonjang commited on
Commit
193b667
·
verified ·
1 Parent(s): e3a7ffa

HTML result tables, KURE rows in crimson

Browse files
Files changed (1) hide show
  1. README.md +117 -61
README.md CHANGED
@@ -137,25 +137,39 @@ The figure plots mean nDCG@10 against mean PPS across the nine Korean benchmarks
137
 
138
  ### Results — MTEB-ko-retrieval (9 subsets)
139
  <!-- **공식 9개 subset 을 모두 평가한 모델**의 9-subset mean NDCG@10, PPS -->
140
- | Model | Params | Mean NDCG@10 | Mean PPS |
141
- |---|---|---|---|
142
- | Qwen/Qwen3-Reranker-8B | 8.2B | 0.8990 | 15.0 |
143
- | Qwen/Qwen3-Reranker-4B | 4.0B | 0.8916 | 24.3 |
144
- | KaLM-Embedding/KaLM-Reranker-V1-Large-R2 | 7.5B | 0.8910 | 25.3 |
145
- | **nlpai-lab/KURE-Reranker-base** | 1.7B | 0.8800 | 55.0 |
146
- | **nlpai-lab/KURE-Reranker-nano** | 149M | 0.8757 | 482.4 |
147
- | zeroentropy/zerank-2-reranker | 4.0B | 0.8652 | 29.3 |
148
- | mixedbread-ai/mxbai-rerank-large-v2 | 1.5B | 0.8607 | 64.8 |
149
- | lightonai/LightOn-rerank-PW-4B | 4.5B | 0.8605 | 15.1 |
150
- | Qwen/Qwen3-Reranker-0.6B | 596M | 0.8528 | 101.0 |
151
- | BAAI/bge-reranker-v2-m3 | 568M | 0.8518 | 412.6 |
152
- | nvidia/llama-nemotron-rerank-1b-v2 | 1.2B | 0.8468 | 128.2 |
153
- | nlpai-lab/LAMAR-600m | 568M | 0.8375 | 415.0 |
154
- | dragonkue/bge-reranker-v2-m3-ko | 568M | 0.8199 | 409.1 |
155
- | BAAI/bge-reranker-v2-gemma | 2.5B | 0.8099 | 58.8 |
156
- | upskyy/ko-reranker-8k | 568M | 0.8054 | 409.8 |
157
- | Dongjin-kr/ko-reranker | 560M | 0.7861 | 488.0 |
158
- | telepix/PIXIE-Spell-Reranker-Preview-0.6B | 596M | 0.7760 | 99.7 |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
159
 
160
  PPS is measured on a single NVIDIA RTX A6000 48GB per model and averaged over the nine benchmarks. For each benchmark, every model is timed on the same ~550 sampled query–document pairs (whole queries, fixed seed). Inputs are length-sorted and dynamically padded within each batch. Models run in bf16 with flash_attention_2, except `KaLM-Embedding/KaLM-Reranker-V1-Large-R2`, which uses sdpa because transformers does not support flash_attention_2 for T5Gemma2. Each query is reranked over its gold documents plus, as negatives, the BM25 top-50 minus the golds; BelebeleRetrieval uses its kor_Hang-kor_Hang subset.
161
 
@@ -164,54 +178,96 @@ PPS is measured on a single NVIDIA RTX A6000 48GB per model and averaged over th
164
  <details>
165
  <summary>Per-dataset NDCG@10</summary>
166
 
167
- | Model | Params | Ko-StrategyQA | AutoRAGRetrieval | PublicHealthQA | BelebeleRetrieval | MIRACLRetrieval | MrTidyRetrieval | MultiLongDocRetrieval | SQuADKorV1Retrieval | LawIRKo |
168
- |---|---|---|---|---|---|---|---|---|---|---|
169
- | Qwen/Qwen3-Reranker-8B | 8.2B | 0.8720 | 0.9653 | 0.8654 | 0.9785 | 0.8513 | 0.8416 | 0.8255 | 0.9897 | 0.9013 |
170
- | Qwen/Qwen3-Reranker-4B | 4.0B | 0.8733 | 0.9630 | 0.8442 | 0.9794 | 0.8559 | 0.8298 | 0.8161 | 0.9861 | 0.8766 |
171
- | KaLM-Embedding/KaLM-Reranker-V1-Large-R2 | 7.5B | 0.8796 | 0.9415 | 0.8633 | 0.9809 | 0.8382 | 0.8500 | 0.7970 | 0.9863 | 0.8824 |
172
- | **nlpai-lab/KURE-Reranker-base** | 1.7B | 0.8548 | 0.9762 | 0.8381 | 0.9767 | 0.8371 | 0.7970 | 0.8163 | 0.9891 | 0.8343 |
173
- | **nlpai-lab/KURE-Reranker-nano** | 149M | 0.8582 | 0.9741 | 0.8078 | 0.9768 | 0.8400 | 0.8011 | 0.7966 | 0.9895 | 0.8368 |
174
- | jinaai/jina-reranker-v3.5 | 597M | 0.8539 | 0.9838 | 0.8337 | 0.9659 | 0.8565 | 0.8194 | — | 0.9887 | 0.8545 |
175
- | jinaai/jina-reranker-v3 | 597M | 0.8553 | 0.9773 | 0.7762 | 0.9606 | 0.8449 | 0.8104 | — | 0.9859 | 0.8546 |
176
- | zeroentropy/zerank-2-reranker | 4.0B | 0.8712 | 0.9436 | 0.8377 | 0.9731 | 0.8003 | 0.8027 | 0.7120 | 0.9791 | 0.8669 |
177
- | mixedbread-ai/mxbai-rerank-large-v2 | 1.5B | 0.8563 | 0.9531 | 0.8407 | 0.9659 | 0.7939 | 0.8771 | 0.6787 | 0.9681 | 0.8130 |
178
- | lightonai/LightOn-rerank-PW-4B | 4.5B | 0.8567 | 0.9321 | 0.8312 | 0.9735 | 0.8072 | 0.8091 | 0.7609 | 0.9803 | 0.7938 |
179
- | Qwen/Qwen3-Reranker-0.6B | 596M | 0.8336 | 0.9321 | 0.8092 | 0.9721 | 0.8470 | 0.7325 | 0.7819 | 0.9804 | 0.7865 |
180
- | BAAI/bge-reranker-v2-m3 | 568M | 0.8487 | 0.9663 | 0.7982 | 0.9725 | 0.8129 | 0.8222 | 0.6690 | 0.9853 | 0.7906 |
181
- | nvidia/llama-nemotron-rerank-1b-v2 | 1.2B | 0.8536 | 0.9480 | 0.8130 | 0.9755 | 0.8182 | 0.8026 | 0.6719 | 0.9858 | 0.7527 |
182
- | nlpai-lab/LAMAR-600m | 568M | 0.8461 | 0.9591 | 0.8044 | 0.9740 | 0.8214 | 0.7975 | 0.5679 | 0.9850 | 0.7822 |
183
- | dragonkue/bge-reranker-v2-m3-ko | 568M | 0.8232 | 0.9684 | 0.8155 | 0.9744 | 0.7573 | 0.6776 | 0.7061 | 0.9846 | 0.6721 |
184
- | BAAI/bge-reranker-v2-gemma | 2.5B | 0.8614 | 0.9407 | 0.8069 | 0.9696 | 0.8362 | 0.8429 | 0.2881 | 0.9858 | 0.7572 |
185
- | upskyy/ko-reranker-8k | 568M | 0.8143 | 0.9230 | 0.7792 | 0.9609 | 0.7249 | 0.6998 | 0.5975 | 0.9718 | 0.7770 |
186
- | Dongjin-kr/ko-reranker | 560M | 0.8468 | 0.9014 | 0.6976 | 0.9655 | 0.8017 | 0.7772 | 0.3721 | 0.9785 | 0.7343 |
187
- | telepix/PIXIE-Spell-Reranker-Preview-0.6B | 596M | 0.8329 | 0.9794 | 0.8211 | 0.9689 | 0.8449 | 0.7650 | 0.1829 | 0.9850 | 0.6042 |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
188
 
189
  </details>
190
 
191
  <details>
192
  <summary>Per-dataset PPS</summary>
193
 
194
- | Model | Params | Ko-StrategyQA | AutoRAGRetrieval | PublicHealthQA | BelebeleRetrieval | MIRACLRetrieval | MrTidyRetrieval | MultiLongDocRetrieval | SQuADKorV1Retrieval | LawIRKo |
195
- |---|---|---|---|---|---|---|---|---|---|---|
196
- | Qwen/Qwen3-Reranker-8B | 8.2B | 16.5 | 7.4 | 15.3 | 20.5 | 25.0 | 26.6 | 0.7 | 10.8 | 12.4 |
197
- | Qwen/Qwen3-Reranker-4B | 4.0B | 26.2 | 11.8 | 25.6 | 34.5 | 40.1 | 42.8 | 1.1 | 17.1 | 19.9 |
198
- | KaLM-Embedding/KaLM-Reranker-V1-Large-R2 | 7.5B | 28.6 | 12.4 | 25.4 | 36.9 | 39.6 | 43.8 | 0.8 | 18.7 | 21.4 |
199
- | **nlpai-lab/KURE-Reranker-base** | 1.7B | 59.2 | 27.3 | 58.0 | 77.0 | 89.1 | 97.1 | 2.6 | 39.1 | 45.7 |
200
- | **nlpai-lab/KURE-Reranker-nano** | 149M | 475.4 | 233.2 | 529.0 | 750.0 | 765.8 | 855.2 | 16.5 | 330.8 | 386.0 |
201
- | jinaai/jina-reranker-v3.5 | 597M | 141.0 | 38.6 | 123.2 | 197.5 | 277.3 | 295.7 | — | 68.4 | 86.7 |
202
- | jinaai/jina-reranker-v3 | 597M | 113.8 | 25.0 | 95.7 | 166.3 | 244.6 | 261.6 | — | 48.7 | 64.0 |
203
- | zeroentropy/zerank-2-reranker | 4.0B | 31.0 | 12.7 | 29.5 | 41.4 | 51.0 | 56.4 | 1.1 | 18.8 | 22.3 |
204
- | mixedbread-ai/mxbai-rerank-large-v2 | 1.5B | 70.0 | 32.9 | 68.0 | 89.8 | 104.2 | 113.6 | 3.2 | 47.0 | 54.7 |
205
- | lightonai/LightOn-rerank-PW-4B | 4.5B | 16.2 | 7.3 | 17.0 | 22.2 | 23.4 | 26.1 | 0.7 | 10.7 | 12.4 |
206
- | Qwen/Qwen3-Reranker-0.6B | 596M | 105.9 | 49.4 | 109.5 | 145.8 | 162.6 | 177.4 | 4.3 | 70.8 | 83.6 |
207
- | BAAI/bge-reranker-v2-m3 | 568M | 420.7 | 195.6 | 440.9 | 651.8 | 653.2 | 724.8 | 9.3 | 271.6 | 345.5 |
208
- | nvidia/llama-nemotron-rerank-1b-v2 | 1.2B | 133.7 | 54.4 | 131.6 | 186.3 | 220.0 | 243.1 | 3.9 | 84.9 | 95.9 |
209
- | nlpai-lab/LAMAR-600m | 568M | 418.0 | 196.5 | 442.5 | 651.1 | 656.7 | 742.5 | 9.2 | 275.0 | 343.8 |
210
- | dragonkue/bge-reranker-v2-m3-ko | 568M | 416.9 | 195.1 | 434.4 | 641.3 | 653.4 | 716.7 | 9.2 | 271.9 | 342.7 |
211
- | BAAI/bge-reranker-v2-gemma | 2.5B | 64.7 | 26.1 | 60.5 | 83.8 | 99.8 | 106.8 | 2.5 | 40.0 | 44.7 |
212
- | upskyy/ko-reranker-8k | 568M | 411.8 | 195.6 | 436.2 | 641.7 | 649.3 | 726.0 | 9.2 | 275.1 | 342.8 |
213
- | Dongjin-kr/ko-reranker | 560M | 520.8 | 258.0 | 468.3 | 644.1 | 747.6 | 765.2 | 272.7 | 334.4 | 381.2 |
214
- | telepix/PIXIE-Spell-Reranker-Preview-0.6B | 596M | 104.6 | 49.4 | 106.3 | 141.5 | 160.5 | 175.6 | 4.3 | 72.1 | 82.6 |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
215
 
216
  Batch size starts at 8 and doubles until an out-of-memory error. Samples are repeated to fill complete batches. After warmup, three full passes are timed with CUDA events. Throughput is the total number of processed pairs divided by the accumulated GPU forward time. The highest-throughput successful batch is reported. Tokenization, data loading, and CPU preprocessing are excluded. The batch-search approach is informed by the [Ettin reranker speed benchmark](https://huggingface.co/blog/ettin-reranker#speed).
217
 
 
137
 
138
  ### Results — MTEB-ko-retrieval (9 subsets)
139
  <!-- **공식 9개 subset 을 모두 평가한 모델**의 9-subset mean NDCG@10, PPS -->
140
+ <div style="width: fit-content; max-width: 100%; margin: 0 auto;">
141
+ <div style="overflow-x: auto;">
142
+ <table cellpadding="6" style="margin: 0 auto; border-collapse: collapse; color: inherit; text-align: right; font-size: 0.85em; line-height: 1.6; font-variant-numeric: tabular-nums;">
143
+ <thead>
144
+ <tr style="border-bottom: 1px solid rgba(128, 128, 128, 0.35);">
145
+ <th scope="col" style="text-align: left;">Model</th>
146
+ <th scope="col">Params</th>
147
+ <th scope="col">Mean NDCG@10</th>
148
+ <th scope="col">Mean PPS</th>
149
+ </tr>
150
+ </thead>
151
+ <tbody>
152
+ <tr><td style="text-align: left; white-space: nowrap;">Qwen/Qwen3-Reranker-8B</td><td>8.2B</td><td>0.8990</td><td>15.0</td></tr>
153
+ <tr><td style="text-align: left; white-space: nowrap;">Qwen/Qwen3-Reranker-4B</td><td>4.0B</td><td>0.8916</td><td>24.3</td></tr>
154
+ <tr><td style="text-align: left; white-space: nowrap;">KaLM-Embedding/KaLM-Reranker-V1-Large-R2</td><td>7.5B</td><td>0.8910</td><td>25.3</td></tr>
155
+ <tr><td style="text-align: left; white-space: nowrap;"><strong style="color: #c9496b;">nlpai-lab/KURE-Reranker-base</strong></td><td><span style="color: #c9496b;">1.7B</span></td><td><span style="color: #c9496b;">0.8800</span></td><td><span style="color: #c9496b;">55.0</span></td></tr>
156
+ <tr><td style="text-align: left; white-space: nowrap;"><strong style="color: #c9496b;">nlpai-lab/KURE-Reranker-nano</strong></td><td><span style="color: #c9496b;">149M</span></td><td><span style="color: #c9496b;">0.8757</span></td><td><span style="color: #c9496b;">482.4</span></td></tr>
157
+ <tr><td style="text-align: left; white-space: nowrap;">zeroentropy/zerank-2-reranker</td><td>4.0B</td><td>0.8652</td><td>29.3</td></tr>
158
+ <tr><td style="text-align: left; white-space: nowrap;">mixedbread-ai/mxbai-rerank-large-v2</td><td>1.5B</td><td>0.8607</td><td>64.8</td></tr>
159
+ <tr><td style="text-align: left; white-space: nowrap;">lightonai/LightOn-rerank-PW-4B</td><td>4.5B</td><td>0.8605</td><td>15.1</td></tr>
160
+ <tr><td style="text-align: left; white-space: nowrap;">Qwen/Qwen3-Reranker-0.6B</td><td>596M</td><td>0.8528</td><td>101.0</td></tr>
161
+ <tr><td style="text-align: left; white-space: nowrap;">BAAI/bge-reranker-v2-m3</td><td>568M</td><td>0.8518</td><td>412.6</td></tr>
162
+ <tr><td style="text-align: left; white-space: nowrap;">nvidia/llama-nemotron-rerank-1b-v2</td><td>1.2B</td><td>0.8468</td><td>128.2</td></tr>
163
+ <tr><td style="text-align: left; white-space: nowrap;">nlpai-lab/LAMAR-600m</td><td>568M</td><td>0.8375</td><td>415.0</td></tr>
164
+ <tr><td style="text-align: left; white-space: nowrap;">dragonkue/bge-reranker-v2-m3-ko</td><td>568M</td><td>0.8199</td><td>409.1</td></tr>
165
+ <tr><td style="text-align: left; white-space: nowrap;">BAAI/bge-reranker-v2-gemma</td><td>2.5B</td><td>0.8099</td><td>58.8</td></tr>
166
+ <tr><td style="text-align: left; white-space: nowrap;">upskyy/ko-reranker-8k</td><td>568M</td><td>0.8054</td><td>409.8</td></tr>
167
+ <tr><td style="text-align: left; white-space: nowrap;">Dongjin-kr/ko-reranker</td><td>560M</td><td>0.7861</td><td>488.0</td></tr>
168
+ <tr><td style="text-align: left; white-space: nowrap;">telepix/PIXIE-Spell-Reranker-Preview-0.6B</td><td>596M</td><td>0.7760</td><td>99.7</td></tr>
169
+ </tbody>
170
+ </table>
171
+ </div>
172
+ </div>
173
 
174
  PPS is measured on a single NVIDIA RTX A6000 48GB per model and averaged over the nine benchmarks. For each benchmark, every model is timed on the same ~550 sampled query–document pairs (whole queries, fixed seed). Inputs are length-sorted and dynamically padded within each batch. Models run in bf16 with flash_attention_2, except `KaLM-Embedding/KaLM-Reranker-V1-Large-R2`, which uses sdpa because transformers does not support flash_attention_2 for T5Gemma2. Each query is reranked over its gold documents plus, as negatives, the BM25 top-50 minus the golds; BelebeleRetrieval uses its kor_Hang-kor_Hang subset.
175
 
 
178
  <details>
179
  <summary>Per-dataset NDCG@10</summary>
180
 
181
+ <div style="width: fit-content; max-width: 100%; margin: 0 auto;">
182
+ <div style="overflow-x: auto;">
183
+ <table cellpadding="6" style="margin: 0 auto; border-collapse: collapse; color: inherit; text-align: right; font-size: 0.85em; line-height: 1.6; font-variant-numeric: tabular-nums;">
184
+ <thead>
185
+ <tr style="border-bottom: 1px solid rgba(128, 128, 128, 0.35);">
186
+ <th scope="col" style="text-align: left;">Model</th>
187
+ <th scope="col">Params</th>
188
+ <th scope="col">Ko-StrategyQA</th>
189
+ <th scope="col">AutoRAGRetrieval</th>
190
+ <th scope="col">PublicHealthQA</th>
191
+ <th scope="col">BelebeleRetrieval</th>
192
+ <th scope="col">MIRACLRetrieval</th>
193
+ <th scope="col">MrTidyRetrieval</th>
194
+ <th scope="col">MultiLongDocRetrieval</th>
195
+ <th scope="col">SQuADKorV1Retrieval</th>
196
+ <th scope="col">LawIRKo</th>
197
+ </tr>
198
+ </thead>
199
+ <tbody>
200
+ <tr><td style="text-align: left; white-space: nowrap;">Qwen/Qwen3-Reranker-8B</td><td>8.2B</td><td>0.8720</td><td>0.9653</td><td>0.8654</td><td>0.9785</td><td>0.8513</td><td>0.8416</td><td>0.8255</td><td>0.9897</td><td>0.9013</td></tr>
201
+ <tr><td style="text-align: left; white-space: nowrap;">Qwen/Qwen3-Reranker-4B</td><td>4.0B</td><td>0.8733</td><td>0.9630</td><td>0.8442</td><td>0.9794</td><td>0.8559</td><td>0.8298</td><td>0.8161</td><td>0.9861</td><td>0.8766</td></tr>
202
+ <tr><td style="text-align: left; white-space: nowrap;">KaLM-Embedding/KaLM-Reranker-V1-Large-R2</td><td>7.5B</td><td>0.8796</td><td>0.9415</td><td>0.8633</td><td>0.9809</td><td>0.8382</td><td>0.8500</td><td>0.7970</td><td>0.9863</td><td>0.8824</td></tr>
203
+ <tr><td style="text-align: left; white-space: nowrap;"><strong style="color: #c9496b;">nlpai-lab/KURE-Reranker-base</strong></td><td><span style="color: #c9496b;">1.7B</span></td><td><span style="color: #c9496b;">0.8548</span></td><td><span style="color: #c9496b;">0.9762</span></td><td><span style="color: #c9496b;">0.8381</span></td><td><span style="color: #c9496b;">0.9767</span></td><td><span style="color: #c9496b;">0.8371</span></td><td><span style="color: #c9496b;">0.7970</span></td><td><span style="color: #c9496b;">0.8163</span></td><td><span style="color: #c9496b;">0.9891</span></td><td><span style="color: #c9496b;">0.8343</span></td></tr>
204
+ <tr><td style="text-align: left; white-space: nowrap;"><strong style="color: #c9496b;">nlpai-lab/KURE-Reranker-nano</strong></td><td><span style="color: #c9496b;">149M</span></td><td><span style="color: #c9496b;">0.8582</span></td><td><span style="color: #c9496b;">0.9741</span></td><td><span style="color: #c9496b;">0.8078</span></td><td><span style="color: #c9496b;">0.9768</span></td><td><span style="color: #c9496b;">0.8400</span></td><td><span style="color: #c9496b;">0.8011</span></td><td><span style="color: #c9496b;">0.7966</span></td><td><span style="color: #c9496b;">0.9895</span></td><td><span style="color: #c9496b;">0.8368</span></td></tr>
205
+ <tr><td style="text-align: left; white-space: nowrap;">jinaai/jina-reranker-v3.5</td><td>597M</td><td>0.8539</td><td>0.9838</td><td>0.8337</td><td>0.9659</td><td>0.8565</td><td>0.8194</td><td>—</td><td>0.9887</td><td>0.8545</td></tr>
206
+ <tr><td style="text-align: left; white-space: nowrap;">jinaai/jina-reranker-v3</td><td>597M</td><td>0.8553</td><td>0.9773</td><td>0.7762</td><td>0.9606</td><td>0.8449</td><td>0.8104</td><td>—</td><td>0.9859</td><td>0.8546</td></tr>
207
+ <tr><td style="text-align: left; white-space: nowrap;">zeroentropy/zerank-2-reranker</td><td>4.0B</td><td>0.8712</td><td>0.9436</td><td>0.8377</td><td>0.9731</td><td>0.8003</td><td>0.8027</td><td>0.7120</td><td>0.9791</td><td>0.8669</td></tr>
208
+ <tr><td style="text-align: left; white-space: nowrap;">mixedbread-ai/mxbai-rerank-large-v2</td><td>1.5B</td><td>0.8563</td><td>0.9531</td><td>0.8407</td><td>0.9659</td><td>0.7939</td><td>0.8771</td><td>0.6787</td><td>0.9681</td><td>0.8130</td></tr>
209
+ <tr><td style="text-align: left; white-space: nowrap;">lightonai/LightOn-rerank-PW-4B</td><td>4.5B</td><td>0.8567</td><td>0.9321</td><td>0.8312</td><td>0.9735</td><td>0.8072</td><td>0.8091</td><td>0.7609</td><td>0.9803</td><td>0.7938</td></tr>
210
+ <tr><td style="text-align: left; white-space: nowrap;">Qwen/Qwen3-Reranker-0.6B</td><td>596M</td><td>0.8336</td><td>0.9321</td><td>0.8092</td><td>0.9721</td><td>0.8470</td><td>0.7325</td><td>0.7819</td><td>0.9804</td><td>0.7865</td></tr>
211
+ <tr><td style="text-align: left; white-space: nowrap;">BAAI/bge-reranker-v2-m3</td><td>568M</td><td>0.8487</td><td>0.9663</td><td>0.7982</td><td>0.9725</td><td>0.8129</td><td>0.8222</td><td>0.6690</td><td>0.9853</td><td>0.7906</td></tr>
212
+ <tr><td style="text-align: left; white-space: nowrap;">nvidia/llama-nemotron-rerank-1b-v2</td><td>1.2B</td><td>0.8536</td><td>0.9480</td><td>0.8130</td><td>0.9755</td><td>0.8182</td><td>0.8026</td><td>0.6719</td><td>0.9858</td><td>0.7527</td></tr>
213
+ <tr><td style="text-align: left; white-space: nowrap;">nlpai-lab/LAMAR-600m</td><td>568M</td><td>0.8461</td><td>0.9591</td><td>0.8044</td><td>0.9740</td><td>0.8214</td><td>0.7975</td><td>0.5679</td><td>0.9850</td><td>0.7822</td></tr>
214
+ <tr><td style="text-align: left; white-space: nowrap;">dragonkue/bge-reranker-v2-m3-ko</td><td>568M</td><td>0.8232</td><td>0.9684</td><td>0.8155</td><td>0.9744</td><td>0.7573</td><td>0.6776</td><td>0.7061</td><td>0.9846</td><td>0.6721</td></tr>
215
+ <tr><td style="text-align: left; white-space: nowrap;">BAAI/bge-reranker-v2-gemma</td><td>2.5B</td><td>0.8614</td><td>0.9407</td><td>0.8069</td><td>0.9696</td><td>0.8362</td><td>0.8429</td><td>0.2881</td><td>0.9858</td><td>0.7572</td></tr>
216
+ <tr><td style="text-align: left; white-space: nowrap;">upskyy/ko-reranker-8k</td><td>568M</td><td>0.8143</td><td>0.9230</td><td>0.7792</td><td>0.9609</td><td>0.7249</td><td>0.6998</td><td>0.5975</td><td>0.9718</td><td>0.7770</td></tr>
217
+ <tr><td style="text-align: left; white-space: nowrap;">Dongjin-kr/ko-reranker</td><td>560M</td><td>0.8468</td><td>0.9014</td><td>0.6976</td><td>0.9655</td><td>0.8017</td><td>0.7772</td><td>0.3721</td><td>0.9785</td><td>0.7343</td></tr>
218
+ <tr><td style="text-align: left; white-space: nowrap;">telepix/PIXIE-Spell-Reranker-Preview-0.6B</td><td>596M</td><td>0.8329</td><td>0.9794</td><td>0.8211</td><td>0.9689</td><td>0.8449</td><td>0.7650</td><td>0.1829</td><td>0.9850</td><td>0.6042</td></tr>
219
+ </tbody>
220
+ </table>
221
+ </div>
222
+ </div>
223
 
224
  </details>
225
 
226
  <details>
227
  <summary>Per-dataset PPS</summary>
228
 
229
+ <div style="width: fit-content; max-width: 100%; margin: 0 auto;">
230
+ <div style="overflow-x: auto;">
231
+ <table cellpadding="6" style="margin: 0 auto; border-collapse: collapse; color: inherit; text-align: right; font-size: 0.85em; line-height: 1.6; font-variant-numeric: tabular-nums;">
232
+ <thead>
233
+ <tr style="border-bottom: 1px solid rgba(128, 128, 128, 0.35);">
234
+ <th scope="col" style="text-align: left;">Model</th>
235
+ <th scope="col">Params</th>
236
+ <th scope="col">Ko-StrategyQA</th>
237
+ <th scope="col">AutoRAGRetrieval</th>
238
+ <th scope="col">PublicHealthQA</th>
239
+ <th scope="col">BelebeleRetrieval</th>
240
+ <th scope="col">MIRACLRetrieval</th>
241
+ <th scope="col">MrTidyRetrieval</th>
242
+ <th scope="col">MultiLongDocRetrieval</th>
243
+ <th scope="col">SQuADKorV1Retrieval</th>
244
+ <th scope="col">LawIRKo</th>
245
+ </tr>
246
+ </thead>
247
+ <tbody>
248
+ <tr><td style="text-align: left; white-space: nowrap;">Qwen/Qwen3-Reranker-8B</td><td>8.2B</td><td>16.5</td><td>7.4</td><td>15.3</td><td>20.5</td><td>25.0</td><td>26.6</td><td>0.7</td><td>10.8</td><td>12.4</td></tr>
249
+ <tr><td style="text-align: left; white-space: nowrap;">Qwen/Qwen3-Reranker-4B</td><td>4.0B</td><td>26.2</td><td>11.8</td><td>25.6</td><td>34.5</td><td>40.1</td><td>42.8</td><td>1.1</td><td>17.1</td><td>19.9</td></tr>
250
+ <tr><td style="text-align: left; white-space: nowrap;">KaLM-Embedding/KaLM-Reranker-V1-Large-R2</td><td>7.5B</td><td>28.6</td><td>12.4</td><td>25.4</td><td>36.9</td><td>39.6</td><td>43.8</td><td>0.8</td><td>18.7</td><td>21.4</td></tr>
251
+ <tr><td style="text-align: left; white-space: nowrap;"><strong style="color: #c9496b;">nlpai-lab/KURE-Reranker-base</strong></td><td><span style="color: #c9496b;">1.7B</span></td><td><span style="color: #c9496b;">59.2</span></td><td><span style="color: #c9496b;">27.3</span></td><td><span style="color: #c9496b;">58.0</span></td><td><span style="color: #c9496b;">77.0</span></td><td><span style="color: #c9496b;">89.1</span></td><td><span style="color: #c9496b;">97.1</span></td><td><span style="color: #c9496b;">2.6</span></td><td><span style="color: #c9496b;">39.1</span></td><td><span style="color: #c9496b;">45.7</span></td></tr>
252
+ <tr><td style="text-align: left; white-space: nowrap;"><strong style="color: #c9496b;">nlpai-lab/KURE-Reranker-nano</strong></td><td><span style="color: #c9496b;">149M</span></td><td><span style="color: #c9496b;">475.4</span></td><td><span style="color: #c9496b;">233.2</span></td><td><span style="color: #c9496b;">529.0</span></td><td><span style="color: #c9496b;">750.0</span></td><td><span style="color: #c9496b;">765.8</span></td><td><span style="color: #c9496b;">855.2</span></td><td><span style="color: #c9496b;">16.5</span></td><td><span style="color: #c9496b;">330.8</span></td><td><span style="color: #c9496b;">386.0</span></td></tr>
253
+ <tr><td style="text-align: left; white-space: nowrap;">jinaai/jina-reranker-v3.5</td><td>597M</td><td>141.0</td><td>38.6</td><td>123.2</td><td>197.5</td><td>277.3</td><td>295.7</td><td>—</td><td>68.4</td><td>86.7</td></tr>
254
+ <tr><td style="text-align: left; white-space: nowrap;">jinaai/jina-reranker-v3</td><td>597M</td><td>113.8</td><td>25.0</td><td>95.7</td><td>166.3</td><td>244.6</td><td>261.6</td><td>—</td><td>48.7</td><td>64.0</td></tr>
255
+ <tr><td style="text-align: left; white-space: nowrap;">zeroentropy/zerank-2-reranker</td><td>4.0B</td><td>31.0</td><td>12.7</td><td>29.5</td><td>41.4</td><td>51.0</td><td>56.4</td><td>1.1</td><td>18.8</td><td>22.3</td></tr>
256
+ <tr><td style="text-align: left; white-space: nowrap;">mixedbread-ai/mxbai-rerank-large-v2</td><td>1.5B</td><td>70.0</td><td>32.9</td><td>68.0</td><td>89.8</td><td>104.2</td><td>113.6</td><td>3.2</td><td>47.0</td><td>54.7</td></tr>
257
+ <tr><td style="text-align: left; white-space: nowrap;">lightonai/LightOn-rerank-PW-4B</td><td>4.5B</td><td>16.2</td><td>7.3</td><td>17.0</td><td>22.2</td><td>23.4</td><td>26.1</td><td>0.7</td><td>10.7</td><td>12.4</td></tr>
258
+ <tr><td style="text-align: left; white-space: nowrap;">Qwen/Qwen3-Reranker-0.6B</td><td>596M</td><td>105.9</td><td>49.4</td><td>109.5</td><td>145.8</td><td>162.6</td><td>177.4</td><td>4.3</td><td>70.8</td><td>83.6</td></tr>
259
+ <tr><td style="text-align: left; white-space: nowrap;">BAAI/bge-reranker-v2-m3</td><td>568M</td><td>420.7</td><td>195.6</td><td>440.9</td><td>651.8</td><td>653.2</td><td>724.8</td><td>9.3</td><td>271.6</td><td>345.5</td></tr>
260
+ <tr><td style="text-align: left; white-space: nowrap;">nvidia/llama-nemotron-rerank-1b-v2</td><td>1.2B</td><td>133.7</td><td>54.4</td><td>131.6</td><td>186.3</td><td>220.0</td><td>243.1</td><td>3.9</td><td>84.9</td><td>95.9</td></tr>
261
+ <tr><td style="text-align: left; white-space: nowrap;">nlpai-lab/LAMAR-600m</td><td>568M</td><td>418.0</td><td>196.5</td><td>442.5</td><td>651.1</td><td>656.7</td><td>742.5</td><td>9.2</td><td>275.0</td><td>343.8</td></tr>
262
+ <tr><td style="text-align: left; white-space: nowrap;">dragonkue/bge-reranker-v2-m3-ko</td><td>568M</td><td>416.9</td><td>195.1</td><td>434.4</td><td>641.3</td><td>653.4</td><td>716.7</td><td>9.2</td><td>271.9</td><td>342.7</td></tr>
263
+ <tr><td style="text-align: left; white-space: nowrap;">BAAI/bge-reranker-v2-gemma</td><td>2.5B</td><td>64.7</td><td>26.1</td><td>60.5</td><td>83.8</td><td>99.8</td><td>106.8</td><td>2.5</td><td>40.0</td><td>44.7</td></tr>
264
+ <tr><td style="text-align: left; white-space: nowrap;">upskyy/ko-reranker-8k</td><td>568M</td><td>411.8</td><td>195.6</td><td>436.2</td><td>641.7</td><td>649.3</td><td>726.0</td><td>9.2</td><td>275.1</td><td>342.8</td></tr>
265
+ <tr><td style="text-align: left; white-space: nowrap;">Dongjin-kr/ko-reranker</td><td>560M</td><td>520.8</td><td>258.0</td><td>468.3</td><td>644.1</td><td>747.6</td><td>765.2</td><td>272.7</td><td>334.4</td><td>381.2</td></tr>
266
+ <tr><td style="text-align: left; white-space: nowrap;">telepix/PIXIE-Spell-Reranker-Preview-0.6B</td><td>596M</td><td>104.6</td><td>49.4</td><td>106.3</td><td>141.5</td><td>160.5</td><td>175.6</td><td>4.3</td><td>72.1</td><td>82.6</td></tr>
267
+ </tbody>
268
+ </table>
269
+ </div>
270
+ </div>
271
 
272
  Batch size starts at 8 and doubles until an out-of-memory error. Samples are repeated to fill complete batches. After warmup, three full passes are timed with CUDA events. Throughput is the total number of processed pairs divided by the accumulated GPU forward time. The highest-throughput successful batch is reported. Tokenization, data loading, and CPU preprocessing are excluded. The batch-search approach is informed by the [Ettin reranker speed benchmark](https://huggingface.co/blog/ettin-reranker#speed).
273