yjoonjang commited on
Commit
a64a617
·
verified ·
1 Parent(s): 5cbe128

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +10 -11
README.md CHANGED
@@ -28,7 +28,6 @@ tags:
28
  - **Strong Korean reranking:** 88.49 average nDCG@10 across nine Korean benchmarks, the highest average among the evaluated models with fewer than 4B parameters and within 1.1 points of Qwen3-Reranker-4B.
29
  - **Long-document reranking:** 81.63 nDCG@10 on MultiLongDocRetrieval (MLDR), ranking second only to Qwen3-Reranker-8B (82.20) among the evaluated models.
30
  - **Efficient:** 55.1 pairs per second averaged over nine benchmarks on an NVIDIA RTX A6000 48GB, about 2.3× Qwen3-Reranker-4B and 3.6× Qwen3-Reranker-8B under the measurement conditions described below.
31
- - **Long-input support:** Up to **8,192 tokens** for the combined query–document input.
32
 
33
  ## Model Overview
34
 
@@ -36,7 +35,7 @@ tags:
36
  |---|---|
37
  | Model type | Cross-encoder reranker (causal LM with yes/no logit scoring) |
38
  | Backbone | Qwen3-1.7B |
39
- | Parameters | Approximately 1.7B |
40
  | Languages | Korean and English |
41
  | Maximum inference input length | 8,192 tokens, including instruction, query, document, and prompt template |
42
  | Output | One scalar relevance score per query–document pair: logit("yes") − logit("no") |
@@ -140,7 +139,7 @@ Evaluation was conducted using the [reranker-simple-benchmark](https://github.co
140
 
141
  <img src="./assets/pps_vs_ndcg9.png" width="700" alt="Reranking performance (mean nDCG@10) vs. throughput (mean PPS)">
142
 
143
- The figure plots mean nDCG@10 against mean PPS across the nine Korean benchmarks, including MLDR. Dot color marks model size. Throughput uses each model's own input limit and its best measured batch size, so it measures neither speed at equal token lengths nor end-to-end retrieval latency. For example, `Dongjin-kr/ko-reranker` is capped at 512 tokens, which inflates its MLDR throughput.
144
 
145
  ### Results — MTEB-ko-retrieval (9 subsets)
146
  <!-- **공식 9개 subset 을 모두 평가한 모델**의 9-subset mean NDCG@10, PPS -->
@@ -148,8 +147,8 @@ The figure plots mean nDCG@10 against mean PPS across the nine Korean benchmarks
148
  |---|---|---|---|
149
  | tomaarsen/Qwen3-Reranker-8B-seq-cls | 7.6B | 0.9004 | 15.1 |
150
  | tomaarsen/Qwen3-Reranker-4B-seq-cls | 4.0B | 0.8956 | 24.3 |
151
- | nlpai-lab/KURE-Reranker-base | 1.7B | 0.8849 | 55.1 |
152
- | nlpai-lab/KURE-Reranker-nano | 149M | 0.8808 | 473.7 |
153
  | zeroentropy/zerank-2-reranker | 4.0B | 0.8695 | 29.4 |
154
  | lightonai/LightOn-rerank-PW-4B | 4.5B | 0.8664 | 15.0 |
155
  | mixedbread-ai/mxbai-rerank-large-v2 | 1.5B | 0.8661 | 65.2 |
@@ -164,7 +163,7 @@ The figure plots mean nDCG@10 against mean PPS across the nine Korean benchmarks
164
  | telepix/PIXIE-Spell-Reranker-Preview-0.6B | 596M | 0.7806 | 99.6 |
165
  | cross-encoder/ettin-reranker-1b-v1 | 1.0B | 0.6901 | 49.7 |
166
 
167
- PPS is measured on a single NVIDIA RTX A6000 48GB per model and averaged over the nine benchmarks. The same samples are used across models. Inputs are length-sorted and dynamically padded within each batch.
168
 
169
  `jinaai/jina-reranker-v3` and `jinaai/jina-reranker-v3.5` are listwise rerankers that cannot be compared under the same 8,192-token condition on MLDR, so they are excluded from the 9-subset table and their MLDR cells are left blank below.
170
 
@@ -174,8 +173,8 @@ PPS is measured on a single NVIDIA RTX A6000 48GB per model and averaged over th
174
  |---|---|---|---|---|---|---|---|---|---|---|
175
  | tomaarsen/Qwen3-Reranker-8B-seq-cls | 7.6B | 0.8679 | 0.9546 | 0.8893 | 0.9907 | 0.8490 | 0.8409 | 0.8220 | 0.9880 | 0.9014 |
176
  | tomaarsen/Qwen3-Reranker-4B-seq-cls | 4.0B | 0.8733 | 0.9707 | 0.8685 | 0.9906 | 0.8533 | 0.8321 | 0.8105 | 0.9861 | 0.8752 |
177
- | nlpai-lab/KURE-Reranker-base | 1.7B | 0.8548 | 0.9762 | 0.8716 | 0.9880 | 0.8371 | 0.7970 | 0.8163 | 0.9891 | 0.8343 |
178
- | nlpai-lab/KURE-Reranker-nano | 149M | 0.8582 | 0.9741 | 0.8475 | 0.9830 | 0.8400 | 0.8011 | 0.7966 | 0.9895 | 0.8368 |
179
  | jinaai/jina-reranker-v3.5 | 597M | 0.8539 | 0.9838 | 0.8094 | 0.9733 | 0.8565 | 0.8194 | — | 0.9887 | 0.8545 |
180
  | jinaai/jina-reranker-v3 | 597M | 0.8553 | 0.9773 | 0.7960 | 0.9695 | 0.8449 | 0.8104 | — | 0.9859 | 0.8546 |
181
  | zeroentropy/zerank-2-reranker | 4.0B | 0.8712 | 0.9436 | 0.8646 | 0.9846 | 0.8003 | 0.8027 | 0.7120 | 0.9791 | 0.8669 |
@@ -198,8 +197,8 @@ PPS is measured on a single NVIDIA RTX A6000 48GB per model and averaged over th
198
  |---|---|---|---|---|---|---|---|---|---|---|
199
  | tomaarsen/Qwen3-Reranker-8B-seq-cls | 7.6B | 16.6 | 7.3 | 17.5 | 19.4 | 24.9 | 26.5 | 0.7 | 10.6 | 12.3 |
200
  | tomaarsen/Qwen3-Reranker-4B-seq-cls | 4.0B | 26.1 | 11.8 | 28.4 | 31.5 | 40.0 | 42.4 | 1.1 | 17.2 | 20.0 |
201
- | nlpai-lab/KURE-Reranker-base | 1.7B | 59.2 | 27.3 | 64.6 | 71.1 | 89.1 | 97.1 | 2.6 | 39.1 | 45.7 |
202
- | nlpai-lab/KURE-Reranker-nano | 149M | 475.4 | 233.2 | 569.7 | 630.9 | 765.8 | 855.2 | 16.5 | 330.8 | 386.0 |
203
  | jinaai/jina-reranker-v3.5 | 597M | 141.0 | 38.6 | 148.6 | 174.6 | 277.3 | 295.7 | — | 68.4 | 86.7 |
204
  | jinaai/jina-reranker-v3 | 597M | 113.8 | 25.0 | 121.8 | 169.2 | 244.6 | 261.6 | — | 48.7 | 64.0 |
205
  | zeroentropy/zerank-2-reranker | 4.0B | 31.0 | 12.7 | 33.8 | 37.8 | 51.0 | 56.4 | 1.1 | 18.8 | 22.3 |
@@ -216,7 +215,7 @@ PPS is measured on a single NVIDIA RTX A6000 48GB per model and averaged over th
216
  | telepix/PIXIE-Spell-Reranker-Preview-0.6B | 596M | 104.6 | 49.4 | 117.8 | 129.8 | 160.5 | 175.6 | 4.3 | 72.1 | 82.6 |
217
  | cross-encoder/ettin-reranker-1b-v1 | 1.0B | 52.4 | 19.9 | 51.4 | 65.6 | 92.7 | 97.2 | 3.8 | 31.2 | 33.4 |
218
 
219
- 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); our aggregation uses total pairs divided by total time rather than the median of three throughput measurements.
220
 
221
  ## Citation
222
 
 
28
  - **Strong Korean reranking:** 88.49 average nDCG@10 across nine Korean benchmarks, the highest average among the evaluated models with fewer than 4B parameters and within 1.1 points of Qwen3-Reranker-4B.
29
  - **Long-document reranking:** 81.63 nDCG@10 on MultiLongDocRetrieval (MLDR), ranking second only to Qwen3-Reranker-8B (82.20) among the evaluated models.
30
  - **Efficient:** 55.1 pairs per second averaged over nine benchmarks on an NVIDIA RTX A6000 48GB, about 2.3× Qwen3-Reranker-4B and 3.6× Qwen3-Reranker-8B under the measurement conditions described below.
 
31
 
32
  ## Model Overview
33
 
 
35
  |---|---|
36
  | Model type | Cross-encoder reranker (causal LM with yes/no logit scoring) |
37
  | Backbone | Qwen3-1.7B |
38
+ | Parameters | 1.7B |
39
  | Languages | Korean and English |
40
  | Maximum inference input length | 8,192 tokens, including instruction, query, document, and prompt template |
41
  | Output | One scalar relevance score per query–document pair: logit("yes") − logit("no") |
 
139
 
140
  <img src="./assets/pps_vs_ndcg9.png" width="700" alt="Reranking performance (mean nDCG@10) vs. throughput (mean PPS)">
141
 
142
+ The figure plots mean nDCG@10 against mean PPS across the nine Korean benchmarks, including MLDR. Dot color marks model size. Throughput uses each model's own input limit and its best measured batch size, so it measures neither speed at equal token lengths nor end-to-end retrieval latency.
143
 
144
  ### Results — MTEB-ko-retrieval (9 subsets)
145
  <!-- **공식 9개 subset 을 모두 평가한 모델**의 9-subset mean NDCG@10, PPS -->
 
147
  |---|---|---|---|
148
  | tomaarsen/Qwen3-Reranker-8B-seq-cls | 7.6B | 0.9004 | 15.1 |
149
  | tomaarsen/Qwen3-Reranker-4B-seq-cls | 4.0B | 0.8956 | 24.3 |
150
+ | **nlpai-lab/KURE-Reranker-base** | 1.7B | 0.8849 | 55.1 |
151
+ | **nlpai-lab/KURE-Reranker-nano** | 149M | 0.8808 | 473.7 |
152
  | zeroentropy/zerank-2-reranker | 4.0B | 0.8695 | 29.4 |
153
  | lightonai/LightOn-rerank-PW-4B | 4.5B | 0.8664 | 15.0 |
154
  | mixedbread-ai/mxbai-rerank-large-v2 | 1.5B | 0.8661 | 65.2 |
 
163
  | telepix/PIXIE-Spell-Reranker-Preview-0.6B | 596M | 0.7806 | 99.6 |
164
  | cross-encoder/ettin-reranker-1b-v1 | 1.0B | 0.6901 | 49.7 |
165
 
166
+ PPS is measured on a single NVIDIA RTX A6000 48GB per model and averaged over the nine benchmarks.Inputs are length-sorted and dynamically padded within each batch.
167
 
168
  `jinaai/jina-reranker-v3` and `jinaai/jina-reranker-v3.5` are listwise rerankers that cannot be compared under the same 8,192-token condition on MLDR, so they are excluded from the 9-subset table and their MLDR cells are left blank below.
169
 
 
173
  |---|---|---|---|---|---|---|---|---|---|---|
174
  | tomaarsen/Qwen3-Reranker-8B-seq-cls | 7.6B | 0.8679 | 0.9546 | 0.8893 | 0.9907 | 0.8490 | 0.8409 | 0.8220 | 0.9880 | 0.9014 |
175
  | tomaarsen/Qwen3-Reranker-4B-seq-cls | 4.0B | 0.8733 | 0.9707 | 0.8685 | 0.9906 | 0.8533 | 0.8321 | 0.8105 | 0.9861 | 0.8752 |
176
+ | **nlpai-lab/KURE-Reranker-base** | 1.7B | 0.8548 | 0.9762 | 0.8716 | 0.9880 | 0.8371 | 0.7970 | 0.8163 | 0.9891 | 0.8343 |
177
+ | **nlpai-lab/KURE-Reranker-nano** | 149M | 0.8582 | 0.9741 | 0.8475 | 0.9830 | 0.8400 | 0.8011 | 0.7966 | 0.9895 | 0.8368 |
178
  | jinaai/jina-reranker-v3.5 | 597M | 0.8539 | 0.9838 | 0.8094 | 0.9733 | 0.8565 | 0.8194 | — | 0.9887 | 0.8545 |
179
  | jinaai/jina-reranker-v3 | 597M | 0.8553 | 0.9773 | 0.7960 | 0.9695 | 0.8449 | 0.8104 | — | 0.9859 | 0.8546 |
180
  | zeroentropy/zerank-2-reranker | 4.0B | 0.8712 | 0.9436 | 0.8646 | 0.9846 | 0.8003 | 0.8027 | 0.7120 | 0.9791 | 0.8669 |
 
197
  |---|---|---|---|---|---|---|---|---|---|---|
198
  | tomaarsen/Qwen3-Reranker-8B-seq-cls | 7.6B | 16.6 | 7.3 | 17.5 | 19.4 | 24.9 | 26.5 | 0.7 | 10.6 | 12.3 |
199
  | tomaarsen/Qwen3-Reranker-4B-seq-cls | 4.0B | 26.1 | 11.8 | 28.4 | 31.5 | 40.0 | 42.4 | 1.1 | 17.2 | 20.0 |
200
+ | **nlpai-lab/KURE-Reranker-base** | 1.7B | 59.2 | 27.3 | 64.6 | 71.1 | 89.1 | 97.1 | 2.6 | 39.1 | 45.7 |
201
+ | **nlpai-lab/KURE-Reranker-nano** | 149M | 475.4 | 233.2 | 569.7 | 630.9 | 765.8 | 855.2 | 16.5 | 330.8 | 386.0 |
202
  | jinaai/jina-reranker-v3.5 | 597M | 141.0 | 38.6 | 148.6 | 174.6 | 277.3 | 295.7 | — | 68.4 | 86.7 |
203
  | jinaai/jina-reranker-v3 | 597M | 113.8 | 25.0 | 121.8 | 169.2 | 244.6 | 261.6 | — | 48.7 | 64.0 |
204
  | zeroentropy/zerank-2-reranker | 4.0B | 31.0 | 12.7 | 33.8 | 37.8 | 51.0 | 56.4 | 1.1 | 18.8 | 22.3 |
 
215
  | telepix/PIXIE-Spell-Reranker-Preview-0.6B | 596M | 104.6 | 49.4 | 117.8 | 129.8 | 160.5 | 175.6 | 4.3 | 72.1 | 82.6 |
216
  | cross-encoder/ettin-reranker-1b-v1 | 1.0B | 52.4 | 19.9 | 51.4 | 65.6 | 92.7 | 97.2 | 3.8 | 31.2 | 33.4 |
217
 
218
+ 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).
219
 
220
  ## Citation
221