File size: 21,338 Bytes
7749d67 17ef67a 7749d67 01e4c53 7749d67 01e4c53 56c82d0 01e4c53 6602205 c74de6a 01e4c53 7749d67 01e4c53 7749d67 01e4c53 7749d67 01e4c53 7749d67 01e4c53 7749d67 01e4c53 7749d67 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 |
---
license: apache-2.0
language:
- ja
pipeline_tag: text-generation
tags:
- RAG
---
# Kurage
<p align="center">
<img src="https://cdn-uploads.huggingface.co/production/uploads/64b63f8ad57e02621dc93c8b/_SkPhhsg40juscfv9dU4v.jpeg" alt="An anime image of a pink and blue jellyfish surrounded by bubbles" width=500 style="border: 5px solid #3d3c3c"/>
</p>
Kurage is a multipurpose RAG model from [Lightblue](https://huggingface.co/lightblue).
This version of the model has been trained to perform RAG in Japanese.
Features of these models include:
* **Multi-chunk RAG** - Performs RAG using multiple contexts at once.
* **Single-chunk RAG** - Performs RAG using one context at a time, allowing for parallel computing.
* **Answer extension** - Prompts the model to write a longer answer to a given question.
* **Multilingual RAG** - Performs RAG using contexts in languages different to the language of the question.
* **Q&A generation** - Generates questions and answers from a reference text in order to pre-index a set of texts.
Find out how to use these features below.
For models in other languages check [our Kurage collection](https://huggingface.co/collections/lightblue/kurage-66e40cbcc3b3a128bdf031f2). A multilingual model is coming soon!
This model was trained using a ml.gu7ef.8xlarge-gu100 instance on [Platform For AI](https://www.alibabacloud.com/en/product/machine-learning) from [Alibaba Cloud](https://www.alibabacloud.com/).
Note - There is a known issue with the single-chunk RAG mode sometimes saying that it cannot answer a question based on the text when it actually can. This was because our single-chunk training data was 50:50 answers vs cannot answer scenarios, making the model overly conservative. We will address this in a week or two when we re-train using 90:10 data with [the coming release of Qwen 2.5](https://x.com/zhouwenmeng/status/1834899729165304198).
# Basic usage
To use the model for basic multi-chunk RAG, you can use the following code:
```python
from vllm import LLM, SamplingParams
llm = LLM(model="lightblue/kurage-ja")
sampling_params = SamplingParams(temperature=0.2, top_p=0.95, max_tokens=128)
def create_rag_prompt(contexts, question):
context_str = "\n\n".join([f"<<Chunk {i+1}>>\n{x}" for i, x in enumerate(contexts)])
str_inputs = f"""{context_str}
<<Question>>
{question}"""
chat = [
{"role": "user", "content": str_inputs},
]
return llm.llm_engine.tokenizer.tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
contexts = [
"日銀の中川順子審議委員は11日、実質金利は現在きわめて低い水準にあるとした上で、先行き日銀の経済・物価見通しが実現していくとすれば、物価目標実現の観点から金融緩和の度合いを調整していくことになると述べた。",
"7月の日本の経常収支は3.2兆円の黒字となり、7月としては過去最高の黒字額を記録した。しかし、黒字に貢献しているのは相変わらず第一次所得収支の黒字で、7月は4.4兆円の黒字を記録し、1カ月の黒字額としては過去最高を記録した。",
"鈴木俊一財務相は10日付で元財務省関税局長の諏訪園健司氏を新しい日銀理事に任命した。9日に任期満了で退任した貝塚正彰前理事の後任で、任期は4年。",
"8月の円高局面で、日本の機関投資家が過去最大の対外証券投資に動いていたことが、外為市場で話題となっている。"
]
question = "現在、日本の第一次所得収支はいくらですか?"
inputs = create_rag_prompt(contexts, question)
outputs = llm.generate([inputs], sampling_params)
print(outputs[0].outputs[0].text)
# <<References>>
# 2
#
# <<Answer>>
# 4.4兆円
```
# Feature: Multi-chunk RAG
This model can take multiple contexts and a question as input, and it will first output the references of the relevant contexts before outputting an answer to the question.
<details>
<summary>Prompt style</summary>
### Input:
```markdown
<<Chunk 1>>
Junko Nakagawa, a member of the Bank of Japan's Policy Board, stated on the 11th that real interest rates are currently at an extremely low level.
She mentioned that if the BOJ's economic and price outlook materializes in the future, the degree of monetary easing would be adjusted from the perspective of achieving the price target.
<<Chunk 2>>
Japan's current account surplus in July was 3.2 trillion yen, the highest monthly surplus on record for the month of July.
However, the surplus continues to be driven by the primary income balance, which recorded a surplus of 4.4 trillion yen in July, the highest monthly figure on record.
<<Chunk 3>>
Finance Minister Shunichi Suzuki appointed Kenji Suwazono, former Director-General of the Customs and Tariff Bureau at the Ministry of Finance, as the new Executive Director of the Bank of Japan effective the 10th. Suwazono succeeds Masaaki Kaizuka, whose term ended on the 9th, and his term will last for four years.
<<Chunk 4>>
In the yen appreciation phase of August, it has become a topic in the foreign exchange market that Japanese institutional investors engaged in the largest-ever outward securities investment.
<<Question>>
What is Japan's primary income balance currently?
```
### Output:
```markdown
<<References>>
2
<<Answer>>
4.4 trillion yen
```
</details>
<details>
<summary>Python code</summary>
```python
from vllm import LLM, SamplingParams
llm = LLM(model="lightblue/kurage-ja")
sampling_params = SamplingParams(temperature=0.2, top_p=0.95, max_tokens=128)
def create_rag_prompt(contexts, question):
context_str = "\n\n".join([f"<<Chunk {i+1}>>\n{x}" for i, x in enumerate(contexts)])
str_inputs = f"""{context_str}
<<Question>>
{question}"""
chat = [
{"role": "user", "content": str_inputs},
]
return llm.llm_engine.tokenizer.tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
contexts = [
"Junko Nakagawa, a member of the Bank of Japan's Policy Board, stated on the 11th that real interest rates are currently at an extremely low level. She mentioned that if the BOJ's economic and price outlook materializes in the future, the degree of monetary easing would be adjusted from the perspective of achieving the price target.",
"Japan's current account surplus in July was 3.2 trillion yen, the highest monthly surplus on record for the month of July. However, the surplus continues to be driven by the primary income balance, which recorded a surplus of 4.4 trillion yen in July, the highest monthly figure on record.",
"Finance Minister Shunichi Suzuki appointed Kenji Suwazono, former Director-General of the Customs and Tariff Bureau at the Ministry of Finance, as the new Executive Director of the Bank of Japan effective the 10th. Suwazono succeeds Masaaki Kaizuka, whose term ended on the 9th, and his term will last for four years.",
"In the yen appreciation phase of August, it has become a topic in the foreign exchange market that Japanese institutional investors engaged in the largest-ever outward securities investment."
]
question = "What is Japan's primary income balance currently?"
inputs = create_rag_prompt(contexts, question)
outputs = llm.generate([inputs], sampling_params)
print(outputs[0].outputs[0].text)
# <<References>>
# 2
#
# <<Answer>>
# 4.4 trillion yen.
```
</details>
<br/>
# Feature: Single-chunk RAG
This model can also take a single context and a question as input, and it will determine whether it can answer the question based on the context, outputting an answer if it can. This allows for parallel computing of multiple contexts at the same time.
<details>
<summary>Prompt style</summary>
### Irrelevant context input:
```markdown
<<Chunk 1>>
Junko Nakagawa, a member of the Bank of Japan's Policy Board, stated on the 11th that real interest rates are currently at an extremely low level.
She mentioned that if the BOJ's economic and price outlook materializes in the future, the degree of monetary easing would be adjusted from the perspective of achieving the price target.
<<Question>>
What is Japan's primary income balance currently?
```
### Irrelevant context output:
```markdown
<<References>>
None
```
### Relevant context input:
```markdown
<<Chunk 1>>
Japan's current account surplus in July was 3.2 trillion yen, the highest monthly surplus on record for the month of July.
However, the surplus continues to be driven by the primary income balance, which recorded a surplus of 4.4 trillion yen in July, the highest monthly figure on record.
<<Question>>
What is Japan's primary income balance currently?
```
### Relevant context output:
```markdown
<<References>>
1
<<Answer>>
4.4 trillion yen
```
</details>
<details>
<summary>Python code</summary>
```python
from vllm import LLM, SamplingParams
llm = LLM(model="lightblue/kurage-ja")
sampling_params = SamplingParams(temperature=0.2, top_p=0.95, max_tokens=128)
def create_rag_prompt(contexts, question):
context_str = "\n\n".join([f"<<Chunk {i+1}>>\n{x}" for i, x in enumerate(contexts)])
str_inputs = f"""{context_str}
<<Question>>
{question}"""
chat = [
{"role": "user", "content": str_inputs},
]
return llm.llm_engine.tokenizer.tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
contexts = [
"Junko Nakagawa, a member of the Bank of Japan's Policy Board, stated on the 11th that real interest rates are currently at an extremely low level. She mentioned that if the BOJ's economic and price outlook materializes in the future, the degree of monetary easing would be adjusted from the perspective of achieving the price target.",
"Japan's current account surplus in July was 3.2 trillion yen, the highest monthly surplus on record for the month of July. However, the surplus continues to be driven by the primary income balance, which recorded a surplus of 4.4 trillion yen in July, the highest monthly figure on record.",
"Finance Minister Shunichi Suzuki appointed Kenji Suwazono, former Director-General of the Customs and Tariff Bureau at the Ministry of Finance, as the new Executive Director of the Bank of Japan effective the 10th. Suwazono succeeds Masaaki Kaizuka, whose term ended on the 9th, and his term will last for four years.",
"In the yen appreciation phase of August, it has become a topic in the foreign exchange market that Japanese institutional investors engaged in the largest-ever outward securities investment."
]
question = "What is Japan's primary income balance currently?"
outputs = llm.generate([create_rag_prompt([x], question) for x in contexts], sampling_params)
print("\n\n".join([f"{i+1}.\n{o.outputs[0].text}" for i, o in enumerate(outputs)]))
# 1.
# <<References>>
# None
# 2.
# <<References>>
# 1
#
# <<Answer>>
# 4.4 trillion yen.
# 3.
# <<References>>
# None
# 4.
# <<References>>
# None
```
</details>
<br/>
# Feature: Answer extension
By default, this model is trained to output the shortest possible answer to a question. However, if you require a longer answer, you can prompt the model to write a longer answer by writing " <<Long>>" after your question.
<details>
<summary>Prompt style</summary>
### Input:
```markdown
<<Chunk 1>>
Japan's current account surplus in July was 3.2 trillion yen, the highest monthly surplus on record for the month of July.
However, the surplus continues to be driven by the primary income balance, which recorded a surplus of 4.4 trillion yen in July, the highest monthly figure on record.
<<Question>>
What is Japan's primary income balance currently? <<Long>>
```
### Relevant context output:
```markdown
<<References>>
1
<<Answer>>
4.4 trillion yen
```
</details>
<details>
<summary>Python code</summary>
```python
from vllm import LLM, SamplingParams
llm = LLM(model="lightblue/kurage-ja")
sampling_params = SamplingParams(temperature=0.2, top_p=0.95, max_tokens=128)
def create_rag_prompt(contexts, question):
context_str = "\n\n".join([f"<<Chunk {i+1}>>\n{x}" for i, x in enumerate(contexts)])
str_inputs = f"""{context_str}
<<Question>>
{question}"""
chat = [
{"role": "user", "content": str_inputs},
]
return llm.llm_engine.tokenizer.tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
contexts = [
"Junko Nakagawa, a member of the Bank of Japan's Policy Board, stated on the 11th that real interest rates are currently at an extremely low level. She mentioned that if the BOJ's economic and price outlook materializes in the future, the degree of monetary easing would be adjusted from the perspective of achieving the price target.",
"Japan's current account surplus in July was 3.2 trillion yen, the highest monthly surplus on record for the month of July. However, the surplus continues to be driven by the primary income balance, which recorded a surplus of 4.4 trillion yen in July, the highest monthly figure on record.",
"Finance Minister Shunichi Suzuki appointed Kenji Suwazono, former Director-General of the Customs and Tariff Bureau at the Ministry of Finance, as the new Executive Director of the Bank of Japan effective the 10th. Suwazono succeeds Masaaki Kaizuka, whose term ended on the 9th, and his term will last for four years.",
"In the yen appreciation phase of August, it has become a topic in the foreign exchange market that Japanese institutional investors engaged in the largest-ever outward securities investment."
]
question = "What is Japan's primary income balance currently? <<Long>>"
inputs = create_rag_prompt(contexts, question)
outputs = llm.generate([inputs], sampling_params)
print(outputs[0].outputs[0].text)
# <<References>>
# 2
#
# <<Answer>>
# Japan's primary income balance recorded a surplus of 4.4 trillion yen in July.
```
</details>
<br/>
# Feature: Multilinguality
We have trained our model to be able to answer questions in Japanese based on texts in other languages too!
(Note - this is still giving variable results depending on the question and the language of the correct reference. Stay tuned for further improvements in the future.)
<details>
<summary>Prompt style</summary>
### Input:
```markdown
<<Chunk 1>>
Junko Nakagawa, a member of the Bank of Japan's Policy Board, stated on the 11th that real interest rates are currently at an extremely low level.
She mentioned that if the BOJ's economic and price outlook materializes in the future, the degree of monetary easing would be adjusted from the perspective of achieving the price target.
<<Chunk 2>>
7月の日本の経常収支は3.2兆円の黒字となり、7月としては過去最高の黒字額を記録した。しかし、黒字に貢献しているのは相変わらず第一次所得収支の黒字で、7月は4.4兆円の黒字を記録し、1カ月の黒字額としては過去最高を記録した。
<<Chunk 3>>
รัฐมนตรีว่าการกระทรวงการคลัง ชุนอิจิ สุซูกิ ได้แต่งตั้ง เค็นจิ สุวาโซโนะ อดีตอธิบดีกรมศุลกากรและภาษีสิ่งนำเข้าแห่งกระทรวงการคลัง เป็นกรรมการบริหารธนาคารแห่งประเทศญี่ปุ่นคนใหม่ มีผลตั้งแต่วันที่ 10 สุวาโซโนะจะมาแทน มาซาอะกิ ไคซูกะ ที่พ้นวาระไปในวันที่ 9 โดยมีวาระ 4 ปี
<<Chunk 4>>
In the yen appreciation phase of August, it has become a topic in the foreign exchange market that Japanese institutional investors engaged in the largest-ever outward securities investment.
<<Question>>
What is Japan's primary income balance currently?
```
### Output:
```markdown
<<References>>
2
<<Answer>>
4.4 trillion yen
```
</details>
<details>
<summary>Python code</summary>
```python
from vllm import LLM, SamplingParams
llm = LLM(model="lightblue/kurage-ja")
sampling_params = SamplingParams(temperature=0.2, top_p=0.95, max_tokens=128)
def create_rag_prompt(contexts, question):
context_str = "\n\n".join([f"<<Chunk {i+1}>>\n{x}" for i, x in enumerate(contexts)])
str_inputs = f"""{context_str}
<<Question>>
{question}"""
chat = [
{"role": "user", "content": str_inputs},
]
return llm.llm_engine.tokenizer.tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
contexts = [
"นากากาวะ จุนโกะ สมาชิกคณะกรรมการนโยบายการเงิน ธนาคารแห่งประเทศญี่ปุ่น กล่าวในวันที่ 11 ว่า อัตราดอกเบี้ยที่แท้จริงอยู่ในระดับต่ำมากในปัจจุบัน และกล่าวว่า หากแนวโน้มเศรษฐกิจและราคาของธนาคารกลางญี่ปุ่นเป็นจริงในอนาคต การผ่อนคลายนโยบายการเงินจะถูกปรับโดยพิจารณาจากการบรรลุเป้าหมายด้านราคา",
"Der Leistungsbilanzüberschuss Japans betrug im Juli 3,2 Billionen Yen, der höchste monatliche Überschuss aller Zeiten für den Monat Juli. Dieser Überschuss wird jedoch weiterhin durch das positive Primäreinkommen unterstützt, das im Juli einen Überschuss von 4,4 Billionen Yen verzeichnete, die höchste monatliche Zahl in der Geschichte.",
"鈴木俊一財務相は10日付で元財務省関税局長の諏訪園健司氏を新しい日銀理事に任命した。9日に任期満了で退任した貝塚正彰前理事の後任で、任期は4年。",
"Lors de la phase d'appréciation du yen en août, il est devenu un sujet dans le marché des changes que les investisseurs institutionnels japonais ont réalisé la plus grande investissement en titres à l'étranger jamais enregistré."
]
question = "What is Japan's primary income balance currently?"
inputs = create_rag_prompt(contexts, question)
outputs = llm.generate([inputs], sampling_params)
print(outputs[0].outputs[0].text)
# <<References>>
# 2
#
# <<Answer>>
# The primary income balance of Japan is currently 4.4 billion yen.
```
</details>
<br/>
# Feature: Q&A generation
This model can also generate questions and answers based on a piece of text. This can be useful for pre-indexing a database or fine-tuning IR models that will then be used for RAG.
<details>
<summary>Prompt style</summary>
### Input:
```markdown
<<Q&A Generation Context>>
Japan's current account surplus in July was 3.2 trillion yen, the highest monthly surplus on record for the month of July.
However, the surplus continues to be driven by the primary income balance, which recorded a surplus of 4.4 trillion yen in July, the highest monthly figure on record.
```
### Output:
```markdown
<<Question>>
What is Japan's current account surplus in July?
<<Answer>>
3.2 trillion yen
```
</details>
<details>
<summary>Python code</summary>
```python
from vllm import LLM, SamplingParams
llm = LLM(model="lightblue/kurage-ja")
sampling_params = SamplingParams(temperature=0.2, top_p=0.95, max_tokens=128)
context = "Finance Minister Shunichi Suzuki appointed Kenji Suwazono, former Director-General of the Customs and Tariff Bureau at the Ministry of Finance, as the new Executive Director of the Bank of Japan effective the 10th. Suwazono succeeds Masaaki Kaizuka, whose term ended on the 9th, and his term will last for four years.",
def create_qagen_prompt(context):
str_inputs = f"""<<Q&A Generation Context>>
{context}"""
chat = [
{"role": "user", "content": str_inputs},
]
return llm.llm_engine.tokenizer.tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
outputs = llm.generate([create_qagen_prompt(context)], sampling_params)
print("\n\n".join([o.outputs[0].text for o in outputs]))
# <<Question>>
# Who was appointed as the new Executive Director of the Bank of Japan by Finance Minister Shunichi Suzuki?
#
# <<Answer>>
# Kenji Suwazono
```
</details>
<br/>
# Training data
We trained on chunks sourced from the documents in [MADLAD-400](https://huggingface.co/datasets/allenai/MADLAD-400) dataset that
had been evaluated to contain a higher amount of educational information according to a state-of-the-art LLM.
We took chunks of size 250 tokens, 500 tokens, and 1000 tokens randomly for each document.
We then used these chunks to generate questions and answers based on this text using a state-of-the-art LLM.
Finally, we selected negatives for each chunk using the similarity from the dense embeddings of the [BAAI/bge-m3](https://huggingface.co/BAAI/bge-m3) model. |