Update evaluation: PublicHealthQA/Belebele re-evaluated on corrected BM25 pools

f3b3eff verified 5 days ago

12 kB

pipeline_tag: text-ranking
library_name: sentence-transformers
license: apache-2.0
language:
  - ko
  - en
base_model:
  - skt/A.X-Encoder-base
tags:
  - sentence-transformers
  - cross-encoder
  - reranker
  - korean
  - english

KURE-Reranker-nano

KURE-Reranker-nano is a lightweight, approximately 150M-parameter Korean-English bilingual reranker. It jointly encodes a query and a candidate document and predicts a scalar relevance score.

Key Characteristics

Strong Korean reranking: 87.57 average nDCG@10 across nine Korean benchmarks, the highest average among the evaluated models with fewer than 1B parameters and only 0.43 points below KURE-Reranker-base (1.7B).
Long-document reranking: 79.66 nDCG@10 on MultiLongDocRetrieval (MLDR), the best among the evaluated models with fewer than 1B parameters.
High throughput: 482.4 pairs per second averaged over nine benchmarks on an NVIDIA RTX A6000 48GB, about 19.8× Qwen3-Reranker-4B under the measurement conditions described below.
Lightweight: Approximately 150M parameters, based on skt/A.X-Encoder-base.

Model Overview

Property	Value
Model type	Cross-encoder reranker
Backbone	A.X-Encoder-base
Parameters	149M
Languages	Korean and English
Maximum inference input length	8,192 tokens, including query, document, and special tokens
Output	One scalar relevance score per query–document pair
License	Apache-2.0

Usage

Install the required libraries:

pip install "sentence-transformers>=5.1.0" "transformers>=4.55.4" torch

Sentence Transformers

from sentence_transformers import CrossEncoder

model = CrossEncoder("nlpai-lab/KURE-Reranker-nano", max_length=8192)

query = "대한민국의 수도는 어디인가요?"
documents = [
    "대한민국의 수도는 서울입니다.",
    "The capital of South Korea is Seoul.",
    "바나나는 열대 지역에서 재배되는 과일입니다.",
]

results = model.rank(query, documents, batch_size=8, return_documents=True)
for result in results:
    print(f"{result['score']:.4f}\t{result['text']}")

For pair scoring without sorting:

pairs = [(query, document) for document in documents]
scores = model.predict(pairs, batch_size=8)

Transformers

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

model_id = "nlpai-lab/KURE-Reranker-nano"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id).to(device)
model.eval()

query = "대한민국의 수도는 어디인가요?"
documents = [
    "대한민국의 수도는 서울입니다.",
    "The capital of South Korea is Seoul.",
    "바나나는 열대 지역에서 재배되는 과일입니다.",
]
inputs = tokenizer(
    [query] * len(documents), documents,
    padding=True, truncation=True, max_length=8192, return_tensors="pt",
).to(device)

with torch.inference_mode():
    scores = model(**inputs).logits.squeeze(-1).float().cpu()

for index in torch.argsort(scores, descending=True).tolist():
    print(f"{scores[index]:.4f}\t{documents[index]}")

Korean Reranking Evaluation

Scores are nDCG@10, and throughput is pairs per second (PPS). Evaluation was conducted using the reranker-simple-benchmark implementation as a reference.

Reranking performance (mean nDCG@10) vs. throughput (mean PPS)

The figure plots mean nDCG@10 against mean PPS across the nine Korean benchmarks, including MLDR. Dot color marks model size; stars mark the KURE-Reranker models. PPS 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. The x-axis is a non-uniform log scale that stretches 20–26 and compresses 26–50 pairs/s so that nearby points stay readable.

Results — MTEB-ko-retrieval (9 subsets)

Model	Params	Mean NDCG@10	Mean PPS
Qwen/Qwen3-Reranker-8B	8.2B	0.8990	15.0
Qwen/Qwen3-Reranker-4B	4.0B	0.8916	24.3
KaLM-Embedding/KaLM-Reranker-V1-Large-R2	7.5B	0.8910	25.3
nlpai-lab/KURE-Reranker-base	1.7B	0.8800	55.0
nlpai-lab/KURE-Reranker-nano	149M	0.8757	482.4
zeroentropy/zerank-2-reranker	4.0B	0.8652	29.3
mixedbread-ai/mxbai-rerank-large-v2	1.5B	0.8607	64.8
lightonai/LightOn-rerank-PW-4B	4.5B	0.8605	15.1
Qwen/Qwen3-Reranker-0.6B	596M	0.8528	101.0
BAAI/bge-reranker-v2-m3	568M	0.8518	412.6
nvidia/llama-nemotron-rerank-1b-v2	1.2B	0.8468	128.2
nlpai-lab/LAMAR-600m	568M	0.8375	415.0
dragonkue/bge-reranker-v2-m3-ko	568M	0.8199	409.1
BAAI/bge-reranker-v2-gemma	2.5B	0.8099	58.8
upskyy/ko-reranker-8k	568M	0.8054	409.8
Dongjin-kr/ko-reranker	560M	0.7861	488.0
telepix/PIXIE-Spell-Reranker-Preview-0.6B	596M	0.7760	99.7

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.

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.

Per-dataset NDCG@10

Model	Params	Ko-StrategyQA	AutoRAGRetrieval	PublicHealthQA	BelebeleRetrieval	MIRACLRetrieval	MrTidyRetrieval	MultiLongDocRetrieval	SQuADKorV1Retrieval	LawIRKo
Qwen/Qwen3-Reranker-8B	8.2B	0.8720	0.9653	0.8654	0.9785	0.8513	0.8416	0.8255	0.9897	0.9013
Qwen/Qwen3-Reranker-4B	4.0B	0.8733	0.9630	0.8442	0.9794	0.8559	0.8298	0.8161	0.9861	0.8766
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
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
nlpai-lab/KURE-Reranker-nano	149M	0.8582	0.9741	0.8078	0.9768	0.8400	0.8011	0.7966	0.9895	0.8368
jinaai/jina-reranker-v3.5	597M	0.8539	0.9838	0.8337	0.9659	0.8565	0.8194	—	0.9887	0.8545
jinaai/jina-reranker-v3	597M	0.8553	0.9773	0.7762	0.9606	0.8449	0.8104	—	0.9859	0.8546
zeroentropy/zerank-2-reranker	4.0B	0.8712	0.9436	0.8377	0.9731	0.8003	0.8027	0.7120	0.9791	0.8669
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
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
Qwen/Qwen3-Reranker-0.6B	596M	0.8336	0.9321	0.8092	0.9721	0.8470	0.7325	0.7819	0.9804	0.7865
BAAI/bge-reranker-v2-m3	568M	0.8487	0.9663	0.7982	0.9725	0.8129	0.8222	0.6690	0.9853	0.7906
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
nlpai-lab/LAMAR-600m	568M	0.8461	0.9591	0.8044	0.9740	0.8214	0.7975	0.5679	0.9850	0.7822
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
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
upskyy/ko-reranker-8k	568M	0.8143	0.9230	0.7792	0.9609	0.7249	0.6998	0.5975	0.9718	0.7770
Dongjin-kr/ko-reranker	560M	0.8468	0.9014	0.6976	0.9655	0.8017	0.7772	0.3721	0.9785	0.7343
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

Per-dataset PPS

Model	Params	Ko-StrategyQA	AutoRAGRetrieval	PublicHealthQA	BelebeleRetrieval	MIRACLRetrieval	MrTidyRetrieval	MultiLongDocRetrieval	SQuADKorV1Retrieval	LawIRKo
Qwen/Qwen3-Reranker-8B	8.2B	16.5	7.4	15.3	20.5	25.0	26.6	0.7	10.8	12.4
Qwen/Qwen3-Reranker-4B	4.0B	26.2	11.8	25.6	34.5	40.1	42.8	1.1	17.1	19.9
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
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
nlpai-lab/KURE-Reranker-nano	149M	475.4	233.2	529.0	750.0	765.8	855.2	16.5	330.8	386.0
jinaai/jina-reranker-v3.5	597M	141.0	38.6	123.2	197.5	277.3	295.7	—	68.4	86.7
jinaai/jina-reranker-v3	597M	113.8	25.0	95.7	166.3	244.6	261.6	—	48.7	64.0
zeroentropy/zerank-2-reranker	4.0B	31.0	12.7	29.5	41.4	51.0	56.4	1.1	18.8	22.3
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
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
Qwen/Qwen3-Reranker-0.6B	596M	105.9	49.4	109.5	145.8	162.6	177.4	4.3	70.8	83.6
BAAI/bge-reranker-v2-m3	568M	420.7	195.6	440.9	651.8	653.2	724.8	9.3	271.6	345.5
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
nlpai-lab/LAMAR-600m	568M	418.0	196.5	442.5	651.1	656.7	742.5	9.2	275.0	343.8
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
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
upskyy/ko-reranker-8k	568M	411.8	195.6	436.2	641.7	649.3	726.0	9.2	275.1	342.8
Dongjin-kr/ko-reranker	560M	520.8	258.0	468.3	644.1	747.6	765.2	272.7	334.4	381.2
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

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.

Citation

@misc{kure-reranker-nano,
  title  = {KURE-Reranker-nano: A Lightweight Korean–English Bilingual Reranking Model},
  author = {Hong, Seongtae and Jang, Youngjoon and Son, Junyoung and Lee, Taemin and Lim, Heuiseok},
  year   = {2026},
  url    = {https://huggingface.co/nlpai-lab/KURE-Reranker-nano},
}