KURE-Reranker-base

KURE-Reranker-base is a Korean-English bilingual reranker. It reads a query and a candidate document together and predicts a scalar relevance score.

Key Characteristics

• Strong Korean reranking: 88.00 average nDCG@10 across nine Korean benchmarks, the highest average among the evaluated models with fewer than 4B parameters and within 1.2 points of Qwen3-Reranker-4B.
• Long-document reranking: 81.63 nDCG@10 on MultiLongDocRetrieval (MLDR), ranking second only to Qwen3-Reranker-8B (82.55) among the evaluated models.
• Efficient: 55.0 pairs per second averaged over nine benchmarks on an NVIDIA RTX A6000 48GB, about 2.3× Qwen3-Reranker-4B and 3.7× Qwen3-Reranker-8B under the measurement conditions described below.

Model Overview

Usage

Install the required libraries:

pip install "sentence-transformers>=5.6.1" torch

The prompt template ships with the model. The default instruction is "Given a web search query, retrieve relevant passages that answer the query", the same instruction used in training.

import torch
from sentence_transformers import CrossEncoder

model = CrossEncoder("nlpai-lab/KURE-Reranker-base", max_length=8192, model_kwargs={"dtype": torch.bfloat16})

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']}")

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

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "nlpai-lab/KURE-Reranker-base"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
tokenizer = AutoTokenizer.from_pretrained(model_id, padding_side="left")
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16).to(device)
model.eval()

max_length = 8192
instruction = "Given a web search query, retrieve relevant passages that answer the query"
yes_id = tokenizer.convert_tokens_to_ids("yes")
no_id = tokenizer.convert_tokens_to_ids("no")

def format_pair(query, document):
    messages = [
        {"role": "system", "content": instruction},
        {"role": "query", "content": query},
        {"role": "document", "content": document},
    ]
    return tokenizer.apply_chat_template(messages, tokenize=False)

def truncate_document(query, document):
    # Truncate the document, not the formatted text, so the closing assistant turn stays intact.
    budget = max_length - len(tokenizer(format_pair(query, ""), add_special_tokens=False).input_ids)
    ids = tokenizer(document, add_special_tokens=False).input_ids
    return document if len(ids) <= budget else tokenizer.decode(ids[:budget])

query = "대한민국의 수도는 어디인가요?"
documents = [
    "대한민국의 수도는 서울입니다.",
    "The capital of South Korea is Seoul.",
    "바나나는 열대 지역에서 재배되는 과일입니다.",
]
texts = [format_pair(query, truncate_document(query, document)) for document in documents]
inputs = tokenizer(texts, padding=True, add_special_tokens=False, return_tensors="pt").to(device)

with torch.inference_mode():
    logits = model(**inputs, logits_to_keep=1).logits[:, -1]
    scores = (logits[:, yes_id] - logits[:, no_id]).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.

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. 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. 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)

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

Per-dataset PPS

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