File size: 40,578 Bytes
63bb3d9 d56f871 37a841c d56f871 63bb3d9 37a841c 63bb3d9 37a841c 63bb3d9 b3e4c8b 63bb3d9 b3e4c8b 63bb3d9 a5659be 63bb3d9 a5659be 63bb3d9 b3e4c8b 63bb3d9 b3e4c8b 63bb3d9 a5659be d56f871 b3e4c8b d56f871 a5659be d56f871 b3e4c8b d56f871 a5659be 63bb3d9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 | ---
tags:
- sentence-transformers
- sparse-encoder
- sparse
- splade
- learned-sparse-retrieval
- feature-extraction
language:
- ko
- en
base_model: omai-research/milco-650m
pipeline_tag: feature-extraction
library_name: sentence-transformers
license: apache-2.0
---
<a href="https://github.com/nlpai-lab/KURE">
<img src="https://huggingface.co/nlpai-lab/KURE-v2/resolve/main/kure_logo.png" width="50%"/>
</a>
# 🔎 KURE-Sparse-base
**KURE-Sparse-base** is a Korean-English bilingual learned sparse retrieval model. It shows state-of-the-art performance among sparse models supporting Korean, averaging **0.7984** nDCG@10 on MTEB(kor, v2) Retrieval. It also outperforms 8B dense models, such as Qwen3-Embedding-8B and comsat-embed-ko-8b-preview.
It is also served through the [KURE API](https://kure.kullm.ai/).
<p align="center">
<img src="assets/kure_ndcg.png" width="100%" alt="Average nDCG@10 on MTEB(kor, v2) Retrieval of the KURE models and baseline retrievers">
</p>
## Key Characteristics
- **Learned Sparse Retrieval (LSR)**: every vector lives in a fixed vocabulary, so documents can be indexed in an inverted
index and matched terms are inspectable.
- **Dual view**: represented with 30,522 English pivot terms (projected from the multilingual encoder) +
250,002 source-language tokens.
- **Long documents**: up to 8,192 tokens.
- **Sparse at both ends**: on average 100 active dimensions per query and 1,897 per document.
## Usage
### Sentence Transformers
```bash
pip install -U sentence-transformers
```
```python
from sentence_transformers import SparseEncoder
model = SparseEncoder("nlpai-lab/KURE-Sparse-base", trust_remote_code=True)
query = "훈민정음은 언제 만들어졌나요?"
documents = [
"세종대왕은 1443년에 훈민정음을 창제하고 1446년에 이를 반포하였다.",
"김치는 배추나 무를 소금에 절인 뒤 고춧가루와 젓갈을 넣어 발효시킨 음식이다.",
"한라산은 해발 1,947m로 남한에서 가장 높은 산이며 제주도 중앙에 자리한다.",
]
query_embeddings = model.encode_query(query)
document_embeddings = model.encode_document(documents)
print(query_embeddings.shape, document_embeddings.shape)
# torch.Size([280524]) torch.Size([3, 280524])
# Dot-product scores (higher is more relevant)
scores = model.similarity(query_embeddings, document_embeddings)
print(scores)
# tensor([[23.9621, 2.0522, 1.0599]])
```
### API
KURE-Sparse-base is served at [kure.kullm.ai](https://kure.kullm.ai/). Sign in with GitHub there to get an API key.
```python
import os
import requests
URL = "https://kure.kullm.ai/api/v1/sparse-embeddings"
HEADERS = {"Authorization": f"Bearer {os.environ['KURE_API_KEY']}"}
query = "훈민정음은 언제 만들어졌나요?"
documents = [
"세종대왕은 1443년에 훈민정음을 창제하고 1446년에 이를 반포하였다.",
"김치는 배추나 무를 소금에 절인 뒤 고춧가루와 젓갈을 넣어 발효시킨 음식이다.",
"한라산은 해발 1,947m로 남한에서 가장 높은 산이며 제주도 중앙에 자리한다.",
]
# input_type adds the query or document markers. return_tokens=True also returns the term of each index
def encode(texts, input_type):
response = requests.post(URL, headers=HEADERS, json={"model": "kure-sparse-base", "input": texts, "input_type": input_type})
response.raise_for_status()
return [dict(zip(e["indices"], e["values"])) for e in response.json()["data"]]
query_embedding = encode(query, "query")[0]
document_embeddings = encode(documents, "document")
# Dot-product scores (higher is more relevant)
scores = [sum(w * doc.get(i, 0.0) for i, w in query_embedding.items()) for doc in document_embeddings]
print([round(s, 4) for s in scores])
# [23.9825, 2.0616, 1.0396]
```
## Evaluation
KURE-Sparse-base is evaluated on the nine **MTEB(kor, v2) Retrieval** tasks. We report nDCG@10.
<div style="width: fit-content; max-width: 100%; margin: 0 auto;">
<div style="overflow-x: auto;">
<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;">
<thead>
<tr style="border-bottom: 1px solid rgba(128, 128, 128, 0.35);">
<th scope="col" style="text-align: left;">Model</th>
<th scope="col" style="text-align: left;">Type</th>
<th scope="col" style="text-align: left;">Params</th>
<th scope="col">Avg</th>
<th scope="col">AutoRAG</th>
<th scope="col">PubHealthQA</th>
<th scope="col">Ko-StrategyQA</th>
<th scope="col">LawIRKo</th>
<th scope="col">SQuADKorV1</th>
<th scope="col">Belebele (ko-ko)</th>
<th scope="col">MrTidy</th>
<th scope="col">MLDR</th>
<th scope="col">MIRACL</th>
</tr>
</thead>
<tbody>
<tr>
<th colspan="13" scope="rowgroup" style="padding: 16px 6px 6px; text-align: left; font-size: 0.8em; letter-spacing: 0.08em; text-transform: uppercase;">KURE Models</th>
</tr>
<tr><td style="text-align: left; white-space: nowrap;"><strong><a href="https://huggingface.co/nlpai-lab/KURE-Sparse-base" style="color: #c9496b;">nlpai-lab/KURE-Sparse-base</a></strong></td><td style="text-align: left; white-space: nowrap;"><span style="color: #c9496b;">Sparse</span></td><td style="text-align: left; white-space: nowrap;"><span style="color: #c9496b;">0.6B</span></td><td><strong style="color: #c9496b;">0.7984</strong></td><td><span style="color: #c9496b;">0.9445</span></td><td><span style="color: #c9496b;">0.8099</span></td><td><span style="color: #c9496b;">0.8188</span></td><td><span style="color: #c9496b;">0.7244</span></td><td><span style="color: #c9496b;">0.9815</span></td><td><span style="color: #c9496b;">0.9703</span></td><td><span style="color: #c9496b;">0.6255</span></td><td><span style="color: #c9496b;">0.5716</span></td><td><span style="color: #c9496b;">0.7395</span></td></tr>
<tr><td style="text-align: left; white-space: nowrap;"><a href="https://huggingface.co/nlpai-lab/KURE-v3" style="color: #c9496b;">nlpai-lab/KURE-v3</a></td><td style="text-align: left; white-space: nowrap;"><span style="color: #c9496b;">Dense</span></td><td style="text-align: left; white-space: nowrap;"><span style="color: #c9496b;">1.1B</span></td><td><span style="color: #c9496b;">0.8027</span></td><td><span style="color: #c9496b;">0.9324</span></td><td><span style="color: #c9496b;">0.8639</span></td><td><span style="color: #c9496b;">0.8192</span></td><td><span style="color: #c9496b;">0.8158</span></td><td><span style="color: #c9496b;">0.9666</span></td><td><span style="color: #c9496b;">0.9673</span></td><td><span style="color: #c9496b;">0.6320</span></td><td><span style="color: #c9496b;">0.4781</span></td><td><span style="color: #c9496b;">0.7489</span></td></tr>
<tr><td style="text-align: left; white-space: nowrap;"><a href="https://huggingface.co/nlpai-lab/KURE-v2" style="color: #c9496b;">nlpai-lab/KURE-v2</a></td><td style="text-align: left; white-space: nowrap;"><span style="color: #c9496b;">Late-interaction</span></td><td style="text-align: left; white-space: nowrap;"><span style="color: #c9496b;">0.2B</span></td><td><span style="color: #c9496b;">0.8160</span></td><td><span style="color: #c9496b;">0.9718</span></td><td><span style="color: #c9496b;">0.8228</span></td><td><span style="color: #c9496b;">0.8070</span></td><td><span style="color: #c9496b;">0.7550</span></td><td><span style="color: #c9496b;">0.9846</span></td><td><span style="color: #c9496b;">0.9660</span></td><td><span style="color: #c9496b;">0.5974</span></td><td><span style="color: #c9496b;">0.7158</span></td><td><span style="color: #c9496b;">0.7237</span></td></tr>
<tr><td style="text-align: left; white-space: nowrap;"><a href="https://huggingface.co/nlpai-lab/KURE-Sparse-nano" style="color: #c9496b;">nlpai-lab/KURE-Sparse-nano</a></td><td style="text-align: left; white-space: nowrap;"><span style="color: #c9496b;">Sparse</span></td><td style="text-align: left; white-space: nowrap;"><span style="color: #c9496b;">0.1B</span></td><td><span style="color: #c9496b;">0.7716</span></td><td><span style="color: #c9496b;">0.9360</span></td><td><span style="color: #c9496b;">0.7711</span></td><td><span style="color: #c9496b;">0.7924</span></td><td><span style="color: #c9496b;">0.7226</span></td><td><span style="color: #c9496b;">0.9774</span></td><td><span style="color: #c9496b;">0.9608</span></td><td><span style="color: #c9496b;">0.5618</span></td><td><span style="color: #c9496b;">0.5133</span></td><td><span style="color: #c9496b;">0.7091</span></td></tr>
<tr><td style="text-align: left; white-space: nowrap;"><a href="https://huggingface.co/nlpai-lab/KURE-v1" style="color: #c9496b;">nlpai-lab/KURE-v1</a></td><td style="text-align: left; white-space: nowrap;"><span style="color: #c9496b;">Dense</span></td><td style="text-align: left; white-space: nowrap;"><span style="color: #c9496b;">0.6B</span></td><td><span style="color: #c9496b;">0.7616</span></td><td><span style="color: #c9496b;">0.8708</span></td><td><span style="color: #c9496b;">0.8193</span></td><td><span style="color: #c9496b;">0.7999</span></td><td><span style="color: #c9496b;">0.7426</span></td><td><span style="color: #c9496b;">0.9357</span></td><td><span style="color: #c9496b;">0.9502</span></td><td><span style="color: #c9496b;">0.5909</span></td><td><span style="color: #c9496b;">0.4637</span></td><td><span style="color: #c9496b;">0.6816</span></td></tr>
<tr><td style="text-align: left; white-space: nowrap;"><a href="https://huggingface.co/nlpai-lab/KoE5" style="color: #c9496b;">nlpai-lab/KoE5</a></td><td style="text-align: left; white-space: nowrap;"><span style="color: #c9496b;">Dense</span></td><td style="text-align: left; white-space: nowrap;"><span style="color: #c9496b;">0.6B</span></td><td><span style="color: #c9496b;">0.7337</span></td><td><span style="color: #c9496b;">0.8434</span></td><td><span style="color: #c9496b;">0.8351</span></td><td><span style="color: #c9496b;">0.8001</span></td><td><span style="color: #c9496b;">0.7756</span></td><td><span style="color: #c9496b;">0.8980</span></td><td><span style="color: #c9496b;">0.9425</span></td><td><span style="color: #c9496b;">0.5841</span></td><td><span style="color: #c9496b;">0.3015</span></td><td><span style="color: #c9496b;">0.6235</span></td></tr>
</tbody>
<tbody style="border-top: 1px solid rgba(128, 128, 128, 0.25);">
<tr>
<th colspan="13" scope="rowgroup" style="padding: 16px 6px 6px; text-align: left; font-size: 0.8em; letter-spacing: 0.08em; text-transform: uppercase;">Dense (Single-Vector) Models</th>
</tr>
<tr><td style="text-align: left; white-space: nowrap;">nvidia/Nemotron-3-Embed-8B-BF16</td><td style="text-align: left; white-space: nowrap;">Dense</td><td style="text-align: left; white-space: nowrap;">8.0B</td><td>0.8127</td><td>0.9311</td><td>0.8898</td><td>0.8405</td><td>0.8481</td><td>0.9650</td><td>0.9714</td><td>0.6540</td><td>0.4785</td><td>0.7363</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">sionic-ai/comsat-embed-ko-8b-preview</td><td style="text-align: left; white-space: nowrap;">Dense</td><td style="text-align: left; white-space: nowrap;">7.6B</td><td>0.7927</td><td>0.8518</td><td>0.8871</td><td>0.8394</td><td>0.8164</td><td>0.9168</td><td>0.9853</td><td>0.6253</td><td>0.5157</td><td>0.6964</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">Qwen/Qwen3-Embedding-8B</td><td style="text-align: left; white-space: nowrap;">Dense</td><td style="text-align: left; white-space: nowrap;">7.6B</td><td>0.7826</td><td>0.8276</td><td>0.8721</td><td>0.8363</td><td>0.8171</td><td>0.9063</td><td>0.9824</td><td>0.6187</td><td>0.5046</td><td>0.6783</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">nvidia/Nemotron-3-Embed-1B-BF16</td><td style="text-align: left; white-space: nowrap;">Dense</td><td style="text-align: left; white-space: nowrap;">1.1B</td><td>0.7817</td><td>0.9074</td><td>0.8642</td><td>0.8136</td><td>0.7758</td><td>0.9505</td><td>0.9629</td><td>0.6084</td><td>0.4202</td><td>0.7325</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">Qwen/Qwen3-Embedding-4B</td><td style="text-align: left; white-space: nowrap;">Dense</td><td style="text-align: left; white-space: nowrap;">4.0B</td><td>0.7737</td><td>0.8431</td><td>0.8693</td><td>0.8270</td><td>0.7769</td><td>0.9044</td><td>0.9522</td><td>0.6076</td><td>0.5022</td><td>0.6803</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">perplexity-ai/pplx-embed-v1-4b</td><td style="text-align: left; white-space: nowrap;">Dense</td><td style="text-align: left; white-space: nowrap;">4.0B</td><td>0.7705</td><td>0.8321</td><td>0.9042</td><td>0.8326</td><td>0.7736</td><td>0.9402</td><td>0.9624</td><td>0.5778</td><td>0.4343</td><td>0.6775</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">microsoft/harrier-oss-v1-27b</td><td style="text-align: left; white-space: nowrap;">Dense</td><td style="text-align: left; white-space: nowrap;">27.0B</td><td>0.7667</td><td>0.8176</td><td>0.8971</td><td>0.8361</td><td>0.8737</td><td>0.9204</td><td>0.9546</td><td>0.5306</td><td>0.4046</td><td>0.6653</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">dragonkue/snowflake-arctic-embed-l-v2.0-ko</td><td style="text-align: left; white-space: nowrap;">Dense</td><td style="text-align: left; white-space: nowrap;">0.6B</td><td>0.7653</td><td>0.9093</td><td>0.8337</td><td>0.8050</td><td>0.7735</td><td>0.9447</td><td>0.9518</td><td>0.5712</td><td>0.4304</td><td>0.6685</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">codefuse-ai/F2LLM-v2-8B</td><td style="text-align: left; white-space: nowrap;">Dense</td><td style="text-align: left; white-space: nowrap;">7.6B</td><td>0.7638</td><td>0.7678</td><td>0.9380</td><td>0.8371</td><td>0.8405</td><td>0.8874</td><td>0.9513</td><td>0.6162</td><td>0.4047</td><td>0.6313</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">telepix/PIXIE-Rune-v1.5</td><td style="text-align: left; white-space: nowrap;">Dense</td><td style="text-align: left; white-space: nowrap;">0.6B</td><td>0.7618</td><td>0.8927</td><td>0.8426</td><td>0.8064</td><td>0.7705</td><td>0.9457</td><td>0.9617</td><td>0.5492</td><td>0.4482</td><td>0.6393</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">nvidia/llama-nemotron-embed-1b-v2</td><td style="text-align: left; white-space: nowrap;">Dense</td><td style="text-align: left; white-space: nowrap;">1.2B</td><td>0.7564</td><td>0.8708</td><td>0.8177</td><td>0.8055</td><td>0.7329</td><td>0.9345</td><td>0.9519</td><td>0.5988</td><td>0.3952</td><td>0.7004</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">dragonkue/BGE-m3-ko</td><td style="text-align: left; white-space: nowrap;">Dense</td><td style="text-align: left; white-space: nowrap;">0.6B</td><td>0.7547</td><td>0.8738</td><td>0.8155</td><td>0.7959</td><td>0.7322</td><td>0.9414</td><td>0.9503</td><td>0.6099</td><td>0.3899</td><td>0.6833</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">BAAI/bge-m3</td><td style="text-align: left; white-space: nowrap;">Dense</td><td style="text-align: left; white-space: nowrap;">0.6B</td><td>0.7509</td><td>0.8301</td><td>0.8041</td><td>0.7941</td><td>0.7174</td><td>0.9038</td><td>0.9315</td><td>0.6471</td><td>0.4287</td><td>0.7015</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">jinaai/jina-embeddings-v5-text-small-retrieval</td><td style="text-align: left; white-space: nowrap;">Dense</td><td style="text-align: left; white-space: nowrap;">0.6B</td><td>0.7419</td><td>0.8409</td><td>0.8664</td><td>0.8096</td><td>0.7483</td><td>0.8926</td><td>0.9346</td><td>0.5779</td><td>0.3315</td><td>0.6752</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">perplexity-ai/pplx-embed-v1-0.6b</td><td style="text-align: left; white-space: nowrap;">Dense</td><td style="text-align: left; white-space: nowrap;">0.6B</td><td>0.7390</td><td>0.8414</td><td>0.8255</td><td>0.8100</td><td>0.7444</td><td>0.9174</td><td>0.9393</td><td>0.5524</td><td>0.3179</td><td>0.7026</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">google/embeddinggemma-300m</td><td style="text-align: left; white-space: nowrap;">Dense</td><td style="text-align: left; white-space: nowrap;">0.3B</td><td>0.7254</td><td>0.8187</td><td>0.8170</td><td>0.8008</td><td>0.6850</td><td>0.9114</td><td>0.9430</td><td>0.5253</td><td>0.3844</td><td>0.6434</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">Qwen/Qwen3-Embedding-0.6B</td><td style="text-align: left; white-space: nowrap;">Dense</td><td style="text-align: left; white-space: nowrap;">0.6B</td><td>0.7140</td><td>0.8338</td><td>0.8107</td><td>0.7586</td><td>0.7526</td><td>0.8490</td><td>0.9124</td><td>0.4870</td><td>0.4199</td><td>0.6023</td></tr>
</tbody>
<tbody style="border-top: 1px solid rgba(128, 128, 128, 0.25);">
<tr>
<th colspan="13" scope="rowgroup" style="padding: 16px 6px 6px; text-align: left; font-size: 0.8em; letter-spacing: 0.08em; text-transform: uppercase;">Late-interaction (Multi-Vector) Models</th>
</tr>
<tr><td style="text-align: left; white-space: nowrap;">lightonai/mLateOn</td><td style="text-align: left; white-space: nowrap;">Late-interaction</td><td style="text-align: left; white-space: nowrap;">0.3B</td><td>0.7906</td><td>0.9392</td><td>0.8061</td><td>0.7905</td><td>0.6431</td><td>0.9803</td><td>0.9608</td><td>0.5817</td><td>0.7005</td><td>0.7135</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">perplexity-ai/pplx-embed-v1-late-0.6b</td><td style="text-align: left; white-space: nowrap;">Late-interaction</td><td style="text-align: left; white-space: nowrap;">0.6B</td><td>0.7381</td><td>0.8557</td><td>0.8089</td><td>0.7973</td><td>0.7285</td><td>0.9696</td><td>0.9548</td><td>0.5400</td><td>0.2816</td><td>0.7064</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">dragonkue/colbert-ko-0.1b</td><td style="text-align: left; white-space: nowrap;">Late-interaction</td><td style="text-align: left; white-space: nowrap;">0.1B</td><td>0.6776</td><td>0.9700</td><td>0.7482</td><td>0.7364</td><td>0.4475</td><td>0.9794</td><td>0.9644</td><td>0.3966</td><td>0.2872</td><td>0.5685</td></tr>
</tbody>
<tbody style="border-top: 1px solid rgba(128, 128, 128, 0.25);">
<tr>
<th colspan="13" scope="rowgroup" style="padding: 16px 6px 6px; text-align: left; font-size: 0.8em; letter-spacing: 0.08em; text-transform: uppercase;">Sparse Models</th>
</tr>
<tr><td style="text-align: left; white-space: nowrap;">omai-research/milco-650m</td><td style="text-align: left; white-space: nowrap;">Sparse</td><td style="text-align: left; white-space: nowrap;">0.6B</td><td>0.7824</td><td>0.9483</td><td>0.8219</td><td>0.8143</td><td>0.6724</td><td>0.9700</td><td>0.9623</td><td>0.6997</td><td>0.4318</td><td>0.7209</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">telepix/PIXIE-Splade-v1.5</td><td style="text-align: left; white-space: nowrap;">Sparse</td><td style="text-align: left; white-space: nowrap;">0.1B</td><td>0.7608</td><td>0.9412</td><td>0.7910</td><td>0.7993</td><td>0.7667</td><td>0.9687</td><td>0.9651</td><td>0.5132</td><td>0.4938</td><td>0.6085</td></tr>
</tbody>
</table>
</div>
</div>
### Performance-Sparsity
<p align="center">
<img src="assets/ndcg_flops.png" width="80%" alt="Average nDCG@10 vs. average FLOPs of sparse retrievers on MTEB(kor, v2) Retrieval">
</p>
Average active dimensions per vector and FLOPs, averaged over the nine tasks.
<div style="width: fit-content; max-width: 100%; margin: 0 auto;">
<div style="overflow-x: auto;">
<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;">
<thead>
<tr style="border-bottom: 1px solid rgba(128, 128, 128, 0.35);">
<th scope="col" style="text-align: left;">Model</th>
<th scope="col" style="text-align: left;">Params</th>
<th scope="col">Query dims</th>
<th scope="col">Document dims</th>
<th scope="col">FLOPs</th>
</tr>
</thead>
<tbody>
<tr><td style="text-align: left; white-space: nowrap;"><strong style="color: #c9496b;">KURE-Sparse-base</strong></td><td style="text-align: left; white-space: nowrap;"><span style="color: #c9496b;">0.6B</span></td><td><span style="color: #c9496b;">100</span></td><td><span style="color: #c9496b;">1,897</span></td><td><span style="color: #c9496b;">36.5</span></td></tr>
<tr><td style="text-align: left; white-space: nowrap;"><strong style="color: #c9496b;">KURE-Sparse-nano</strong></td><td style="text-align: left; white-space: nowrap;"><span style="color: #c9496b;">0.1B</span></td><td><span style="color: #c9496b;">147</span></td><td><span style="color: #c9496b;">843</span></td><td><span style="color: #c9496b;">23.0</span></td></tr>
<tr><td style="text-align: left; white-space: nowrap;">omai-research/milco-650m</td><td style="text-align: left; white-space: nowrap;">0.6B</td><td>105</td><td>1,167</td><td>32.3</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">telepix/PIXIE-Splade-v1.5</td><td style="text-align: left; white-space: nowrap;">0.1B</td><td>117</td><td>1,541</td><td>37.3</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">yjoonjang/splade-ko-v1</td><td style="text-align: left; white-space: nowrap;">0.1B</td><td>109</td><td>742</td><td>29.1</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">telepix/PIXIE-Splade-v1.0</td><td style="text-align: left; white-space: nowrap;">0.1B</td><td>121</td><td>625</td><td>22.4</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">opensearch-project/opensearch-neural-sparse-encoding-multilingual-v1</td><td style="text-align: left; white-space: nowrap;">0.2B</td><td>29</td><td>176</td><td>6.8</td></tr>
</tbody>
</table>
</div>
</div>
### English Retrieval (BEIR, RTEB)
KURE-Sparse-base is also evaluated on two English benchmarks with nDCG@10: the 13 BEIR tasks (BEIR-13), and the **eight public English** tasks of [RTEB](https://huggingface.co/blog/rteb) outside the code domain. SparseUp runs at its own length limits (128 query and 512 document tokens), and the other models at up to 8,192 tokens. The RTEB average differs from the RTEB leaderboard score, which also covers private and code tasks.
<p align="center">
<img src="assets/beir_rteb_sparse.png" width="100%" alt="Average nDCG@10 on BEIR-13 and on the public RTEB English tasks of the KURE-Sparse models and sparse baselines">
</p>
#### BEIR-13
<div style="width: fit-content; max-width: 100%; margin: 0 auto;">
<div style="overflow-x: auto;">
<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;">
<thead>
<tr style="border-bottom: 1px solid rgba(128, 128, 128, 0.35);">
<th scope="col" style="text-align: left;">Model</th>
<th scope="col" style="text-align: left;">Params</th>
<th scope="col">BEIR-13</th>
<th scope="col">FLOPs</th>
</tr>
</thead>
<tbody>
<tr>
<th colspan="4" scope="rowgroup" style="padding: 16px 6px 6px; text-align: left; font-size: 0.8em; letter-spacing: 0.08em; text-transform: uppercase;">KURE Models</th>
</tr>
<tr><td style="text-align: left; white-space: nowrap;"><strong style="color: #c9496b;">KURE-Sparse-base</strong></td><td style="text-align: left; white-space: nowrap;"><span style="color: #c9496b;">0.6B</span></td><td><strong style="color: #c9496b;">0.5864</strong></td><td><span style="color: #c9496b;">55.8</span></td></tr>
<tr><td style="text-align: left; white-space: nowrap;"><strong style="color: #c9496b;">KURE-Sparse-nano</strong></td><td style="text-align: left; white-space: nowrap;"><span style="color: #c9496b;">0.1B</span></td><td><span style="color: #c9496b;">0.5317</span></td><td><span style="color: #c9496b;">37.0</span></td></tr>
</tbody>
<tbody style="border-top: 1px solid rgba(128, 128, 128, 0.25);">
<tr>
<th colspan="4" scope="rowgroup" style="padding: 16px 6px 6px; text-align: left; font-size: 0.8em; letter-spacing: 0.08em; text-transform: uppercase;">Sparse Models</th>
</tr>
<tr><td style="text-align: left; white-space: nowrap;">utahnlp/laconic-1b</td><td style="text-align: left; white-space: nowrap;">1.2B</td><td>0.5831</td><td>52.6</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">Linkup-Platform/linkup-sparseup-embed-v1</td><td style="text-align: left; white-space: nowrap;">0.2B</td><td>0.5645</td><td>8.5</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">omai-research/milco-650m</td><td style="text-align: left; white-space: nowrap;">0.6B</td><td>0.5545</td><td>20.6</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">opensearch-project/opensearch-neural-sparse-encoding-doc-v3-gte</td><td style="text-align: left; white-space: nowrap;">0.1B</td><td>0.5459</td><td>1.7</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">naver/splade-v3</td><td style="text-align: left; white-space: nowrap;">0.1B</td><td>0.5202</td><td>8.0</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">telepix/PIXIE-Splade-v1.5</td><td style="text-align: left; white-space: nowrap;">0.1B</td><td>0.4149</td><td>125.6</td></tr>
</tbody>
</table>
</div>
</div>
<details>
<summary>Per-task nDCG@10</summary>
<div style="width: fit-content; max-width: 100%; margin: 0 auto;">
<div style="overflow-x: auto;">
<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;">
<thead>
<tr style="border-bottom: 1px solid rgba(128, 128, 128, 0.35);">
<th scope="col" style="text-align: left;">Model</th>
<th scope="col">Avg</th>
<th scope="col">ArguAna</th>
<th scope="col">ClimateFEVER</th>
<th scope="col">DBPedia</th>
<th scope="col">FEVER</th>
<th scope="col">FiQA</th>
<th scope="col">HotpotQA</th>
<th scope="col">NFCorpus</th>
<th scope="col">NQ</th>
<th scope="col">Quora</th>
<th scope="col">SCIDOCS</th>
<th scope="col">SciFact</th>
<th scope="col">TREC-COVID</th>
<th scope="col">Touche-2020</th>
</tr>
</thead>
<tbody>
<tr>
<th colspan="15" scope="rowgroup" style="padding: 16px 6px 6px; text-align: left; font-size: 0.8em; letter-spacing: 0.08em; text-transform: uppercase;">KURE Models</th>
</tr>
<tr><td style="text-align: left; white-space: nowrap;"><strong style="color: #c9496b;">KURE-Sparse-base</strong></td><td><strong style="color: #c9496b;">0.5864</strong></td><td><span style="color: #c9496b;">0.5896</span></td><td><span style="color: #c9496b;">0.3959</span></td><td><span style="color: #c9496b;">0.4837</span></td><td><span style="color: #c9496b;">0.9325</span></td><td><span style="color: #c9496b;">0.4720</span></td><td><span style="color: #c9496b;">0.8084</span></td><td><span style="color: #c9496b;">0.3678</span></td><td><span style="color: #c9496b;">0.6758</span></td><td><span style="color: #c9496b;">0.8252</span></td><td><span style="color: #c9496b;">0.1741</span></td><td><span style="color: #c9496b;">0.7183</span></td><td><span style="color: #c9496b;">0.8260</span></td><td><span style="color: #c9496b;">0.3535</span></td></tr>
<tr><td style="text-align: left; white-space: nowrap;"><strong style="color: #c9496b;">KURE-Sparse-nano</strong></td><td><span style="color: #c9496b;">0.5317</span></td><td><span style="color: #c9496b;">0.5075</span></td><td><span style="color: #c9496b;">0.2865</span></td><td><span style="color: #c9496b;">0.4214</span></td><td><span style="color: #c9496b;">0.9079</span></td><td><span style="color: #c9496b;">0.4080</span></td><td><span style="color: #c9496b;">0.7476</span></td><td><span style="color: #c9496b;">0.3340</span></td><td><span style="color: #c9496b;">0.5647</span></td><td><span style="color: #c9496b;">0.7825</span></td><td><span style="color: #c9496b;">0.1581</span></td><td><span style="color: #c9496b;">0.7120</span></td><td><span style="color: #c9496b;">0.7637</span></td><td><span style="color: #c9496b;">0.3183</span></td></tr>
</tbody>
<tbody style="border-top: 1px solid rgba(128, 128, 128, 0.25);">
<tr>
<th colspan="15" scope="rowgroup" style="padding: 16px 6px 6px; text-align: left; font-size: 0.8em; letter-spacing: 0.08em; text-transform: uppercase;">Sparse Models</th>
</tr>
<tr><td style="text-align: left; white-space: nowrap;">utahnlp/laconic-1b</td><td>0.5831</td><td>0.6127</td><td>0.3952</td><td>0.4705</td><td>0.8871</td><td>0.4427</td><td>0.8018</td><td>0.3958</td><td>0.6642</td><td>0.8072</td><td>0.2397</td><td>0.7596</td><td>0.8300</td><td>0.2737</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">Linkup-Platform/linkup-sparseup-embed-v1</td><td>0.5645</td><td>0.5648</td><td>0.3747</td><td>0.4150</td><td>0.8969</td><td>0.4687</td><td>0.7761</td><td>0.3523</td><td>0.5717</td><td>0.8656</td><td>0.1852</td><td>0.7376</td><td>0.7823</td><td>0.3478</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">omai-research/milco-650m</td><td>0.5545</td><td>0.6157</td><td>0.3079</td><td>0.4501</td><td>0.8353</td><td>0.4262</td><td>0.7780</td><td>0.3629</td><td>0.6495</td><td>0.8818</td><td>0.1704</td><td>0.7020</td><td>0.7476</td><td>0.2813</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">opensearch-project/opensearch-neural-sparse-encoding-doc-v3-gte</td><td>0.5459</td><td>0.5196</td><td>0.3123</td><td>0.4553</td><td>0.8598</td><td>0.4069</td><td>0.7158</td><td>0.3592</td><td>0.5817</td><td>0.8734</td><td>0.1675</td><td>0.7251</td><td>0.7330</td><td>0.3866</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">naver/splade-v3</td><td>0.5202</td><td>0.4880</td><td>0.2566</td><td>0.4514</td><td>0.8098</td><td>0.3810</td><td>0.6924</td><td>0.3630</td><td>0.5867</td><td>0.8138</td><td>0.1557</td><td>0.7171</td><td>0.7346</td><td>0.3119</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">telepix/PIXIE-Splade-v1.5</td><td>0.4149</td><td>0.2185</td><td>0.1302</td><td>0.3519</td><td>0.6475</td><td>0.2753</td><td>0.5756</td><td>0.3325</td><td>0.3858</td><td>0.6878</td><td>0.1468</td><td>0.6694</td><td>0.7068</td><td>0.2650</td></tr>
</tbody>
</table>
</div>
</div>
</details>
#### RTEB
<div style="width: fit-content; max-width: 100%; margin: 0 auto;">
<div style="overflow-x: auto;">
<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;">
<thead>
<tr style="border-bottom: 1px solid rgba(128, 128, 128, 0.35);">
<th scope="col" style="text-align: left;">Model</th>
<th scope="col" style="text-align: left;">Params</th>
<th scope="col">Avg</th>
<th scope="col">FLOPs</th>
</tr>
</thead>
<tbody>
<tr>
<th colspan="4" scope="rowgroup" style="padding: 16px 6px 6px; text-align: left; font-size: 0.8em; letter-spacing: 0.08em; text-transform: uppercase;">KURE Models</th>
</tr>
<tr><td style="text-align: left; white-space: nowrap;"><strong><a href="https://huggingface.co/nlpai-lab/KURE-Sparse-base" style="color: #c9496b;">nlpai-lab/KURE-Sparse-base</a></strong></td><td style="text-align: left; white-space: nowrap;"><span style="color: #c9496b;">0.6B</span></td><td><strong style="color: #c9496b;">0.5093</strong></td><td><span style="color: #c9496b;">277.6</span></td></tr>
<tr><td style="text-align: left; white-space: nowrap;"><a href="https://huggingface.co/nlpai-lab/KURE-Sparse-nano" style="color: #c9496b;">nlpai-lab/KURE-Sparse-nano</a></td><td style="text-align: left; white-space: nowrap;"><span style="color: #c9496b;">0.1B</span></td><td><span style="color: #c9496b;">0.4750</span></td><td><span style="color: #c9496b;">109.9</span></td></tr>
</tbody>
<tbody style="border-top: 1px solid rgba(128, 128, 128, 0.25);">
<tr>
<th colspan="4" scope="rowgroup" style="padding: 16px 6px 6px; text-align: left; font-size: 0.8em; letter-spacing: 0.08em; text-transform: uppercase;">Sparse Models</th>
</tr>
<tr><td style="text-align: left; white-space: nowrap;">utahnlp/laconic-1b</td><td style="text-align: left; white-space: nowrap;">1.2B</td><td>0.5415</td><td>107.2</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">Linkup-Platform/linkup-sparseup-embed-v1</td><td style="text-align: left; white-space: nowrap;">0.2B</td><td>0.4941</td><td>25.1</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">opensearch-project/opensearch-neural-sparse-encoding-doc-v3-gte</td><td style="text-align: left; white-space: nowrap;">0.1B</td><td>0.4721</td><td>12.7</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">naver/splade-v3</td><td style="text-align: left; white-space: nowrap;">0.1B</td><td>0.4596</td><td>35.0</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">omai-research/milco-650m</td><td style="text-align: left; white-space: nowrap;">0.6B</td><td>0.4301</td><td>163.7</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">telepix/PIXIE-Splade-v1.5</td><td style="text-align: left; white-space: nowrap;">0.1B</td><td>0.3630</td><td>391.3</td></tr>
</tbody>
</table>
</div>
</div>
<details>
<summary>Per-task nDCG@10</summary>
<div style="width: fit-content; max-width: 100%; margin: 0 auto;">
<div style="overflow-x: auto;">
<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;">
<thead>
<tr style="border-bottom: 1px solid rgba(128, 128, 128, 0.35);">
<th scope="col" style="text-align: left;">Model</th>
<th scope="col" style="text-align: left;">Params</th>
<th scope="col">Avg</th>
<th scope="col">CUREv1</th>
<th scope="col">ChatDoctor</th>
<th scope="col">FinQA</th>
<th scope="col">AILACasedocs</th>
<th scope="col">LegalSummarization</th>
<th scope="col">HC3Finance</th>
<th scope="col">FinanceBench</th>
<th scope="col">AILAStatutes</th>
<th scope="col">FLOPs</th>
</tr>
</thead>
<tbody>
<tr>
<th colspan="12" scope="rowgroup" style="padding: 16px 6px 6px; text-align: left; font-size: 0.8em; letter-spacing: 0.08em; text-transform: uppercase;">KURE Models</th>
</tr>
<tr><td style="text-align: left; white-space: nowrap;"><strong><a href="https://huggingface.co/nlpai-lab/KURE-Sparse-base" style="color: #c9496b;">nlpai-lab/KURE-Sparse-base</a></strong></td><td style="text-align: left; white-space: nowrap;"><span style="color: #c9496b;">0.6B</span></td><td><strong style="color: #c9496b;">0.5093</strong></td><td><span style="color: #c9496b;">0.5996</span></td><td><span style="color: #c9496b;">0.3914</span></td><td><span style="color: #c9496b;">0.7838</span></td><td><span style="color: #c9496b;">0.2298</span></td><td><span style="color: #c9496b;">0.5494</span></td><td><span style="color: #c9496b;">0.5951</span></td><td><span style="color: #c9496b;">0.7525</span></td><td><span style="color: #c9496b;">0.1727</span></td><td><span style="color: #c9496b;">277.6</span></td></tr>
<tr><td style="text-align: left; white-space: nowrap;"><a href="https://huggingface.co/nlpai-lab/KURE-Sparse-nano" style="color: #c9496b;">nlpai-lab/KURE-Sparse-nano</a></td><td style="text-align: left; white-space: nowrap;"><span style="color: #c9496b;">0.1B</span></td><td><span style="color: #c9496b;">0.4750</span></td><td><span style="color: #c9496b;">0.3591</span></td><td><span style="color: #c9496b;">0.4730</span></td><td><span style="color: #c9496b;">0.7042</span></td><td><span style="color: #c9496b;">0.2860</span></td><td><span style="color: #c9496b;">0.6399</span></td><td><span style="color: #c9496b;">0.4438</span></td><td><span style="color: #c9496b;">0.7005</span></td><td><span style="color: #c9496b;">0.1939</span></td><td><span style="color: #c9496b;">109.9</span></td></tr>
</tbody>
<tbody style="border-top: 1px solid rgba(128, 128, 128, 0.25);">
<tr>
<th colspan="12" scope="rowgroup" style="padding: 16px 6px 6px; text-align: left; font-size: 0.8em; letter-spacing: 0.08em; text-transform: uppercase;">Sparse Models</th>
</tr>
<tr><td style="text-align: left; white-space: nowrap;">utahnlp/laconic-1b</td><td style="text-align: left; white-space: nowrap;">1.2B</td><td>0.5415</td><td>0.5340</td><td>0.5272</td><td>0.6727</td><td>0.3436</td><td>0.6064</td><td>0.5893</td><td>0.7449</td><td>0.3142</td><td>107.2</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">Linkup-Platform/linkup-sparseup-embed-v1</td><td style="text-align: left; white-space: nowrap;">0.2B</td><td>0.4941</td><td>0.4010</td><td>0.5424</td><td>0.5805</td><td>0.2406</td><td>0.6430</td><td>0.5900</td><td>0.6852</td><td>0.2700</td><td>25.1</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">opensearch-project/opensearch-neural-sparse-encoding-doc-v3-gte</td><td style="text-align: left; white-space: nowrap;">0.1B</td><td>0.4721</td><td>0.3004</td><td>0.5015</td><td>0.7344</td><td>0.2596</td><td>0.6554</td><td>0.4265</td><td>0.6855</td><td>0.2139</td><td>12.7</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">naver/splade-v3</td><td style="text-align: left; white-space: nowrap;">0.1B</td><td>0.4596</td><td>0.3306</td><td>0.5036</td><td>0.6374</td><td>0.3153</td><td>0.5663</td><td>0.4118</td><td>0.6878</td><td>0.2241</td><td>35.0</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">omai-research/milco-650m</td><td style="text-align: left; white-space: nowrap;">0.6B</td><td>0.4301</td><td>0.5674</td><td>0.2344</td><td>0.7399</td><td>0.2197</td><td>0.3078</td><td>0.5451</td><td>0.6513</td><td>0.1747</td><td>163.7</td></tr>
<tr><td style="text-align: left; white-space: nowrap;">telepix/PIXIE-Splade-v1.5</td><td style="text-align: left; white-space: nowrap;">0.1B</td><td>0.3630</td><td>0.3414</td><td>0.2178</td><td>0.5751</td><td>0.2082</td><td>0.5216</td><td>0.3042</td><td>0.5180</td><td>0.2181</td><td>391.3</td></tr>
</tbody>
</table>
</div>
</div>
</details>
## Citation
```bibtex
@misc{kure-sparse,
title = {KURE-Sparse: a Korean-English bilingual learned sparse retriever},
author = {Jang, Youngjoon and Son, Junyoung and Lee, Taemin and Hong, Seongtae and Lim, Heuiseok},
year = {2026},
url = {https://huggingface.co/nlpai-lab/KURE-Sparse-base},
}
```
```bibtex
@inproceedings{jang2025kure,
title={KURE: Embedding Model for Korean-Specific Retrieval},
author={Jang, Youngjoon and Son, Junyoung and Lee, Taemin and Hong, Seongtae and Park, JeongBae and Lim, Heuiseok},
booktitle={Annual Conference on Human and Language Technology},
pages={129--134},
year={2025},
organization={Human and Language Technology}
}
```
```bibtex
@inproceedings{nguyen2026milco,
title={MILCO: Learned Sparse Retrieval Across Languages via a Multilingual Connector},
author={Nguyen, Thong and Lei, Yibin and Ju, Jia-Huei and Yang, Eugene and Yates, Andrew},
booktitle={International Conference on Learning Representations},
year={2026}
}
```
|