Kurage

An anime image of a pink and blue jellyfish surrounded by bubbles

Kurage is a multipurpose RAG model from Lightblue based on the Qwen 2 model (Qwen/Qwen2-7B-Instruct).

This version of the model has been trained to perform RAG in 44 languages.

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.

This model was trained using a ml.gu7ef.8xlarge-gu100 instance on Platform For AI from Alibaba Cloud.

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.

Basic usage

To use the model for basic multi-chunk RAG, you can use the following code:

from vllm import LLM, SamplingParams

llm = LLM(model="lightblue/kurage-multilingual")
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.

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.

Prompt style

Input:

<<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:

<<References>>
2

<<Answer>>
4.4 trillion yen
Python code
from vllm import LLM, SamplingParams

llm = LLM(model="lightblue/kurage-multilingual")
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.

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.

Prompt style

Irrelevant context input:

<<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:

<<References>>
None

Relevant context input:

<<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:

<<References>>
1

<<Answer>>
4.4 trillion yen
Python code
from vllm import LLM, SamplingParams

llm = LLM(model="lightblue/kurage-multilingual")
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

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 " <>" after your question.

Prompt style

Input:

<<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:

<<References>>
1

<<Answer>>
4.4 trillion yen
Python code
from vllm import LLM, SamplingParams

llm = LLM(model="lightblue/kurage-multilingual")
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.

Feature: Multilinguality

We have trained our model to be able to answer questions in a given language 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.)

Prompt style

Input:

<<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>>
๏ผ—ๆœˆใฎๆ—ฅๆœฌใฎ็ตŒๅธธๅŽๆ”ฏใฏ๏ผ“๏ผŽ๏ผ’ๅ…†ๅ††ใฎ้ป’ๅญ—ใจใชใ‚Šใ€๏ผ—ๆœˆใจใ—ใฆใฏ้ŽๅŽปๆœ€้ซ˜ใฎ้ป’ๅญ—้กใ‚’่จ˜้Œฒใ—ใŸใ€‚ใ—ใ‹ใ—ใ€้ป’ๅญ—ใซ่ฒข็Œฎใ—ใฆใ„ใ‚‹ใฎใฏ็›ธๅค‰ใ‚ใ‚‰ใš็ฌฌไธ€ๆฌกๆ‰€ๅพ—ๅŽๆ”ฏใฎ้ป’ๅญ—ใงใ€๏ผ—ๆœˆใฏ๏ผ”๏ผŽ๏ผ”ๅ…†ๅ††ใฎ้ป’ๅญ—ใ‚’่จ˜้Œฒใ—ใ€๏ผ‘ใ‚ซๆœˆใฎ้ป’ๅญ—้กใจใ—ใฆใฏ้ŽๅŽปๆœ€้ซ˜ใ‚’่จ˜้Œฒใ—ใŸใ€‚

<<Chunk 3>>
เธฃเธฑเธเธกเธ™เธ•เธฃเธตเธงเนˆเธฒเธเธฒเธฃเธเธฃเธฐเธ—เธฃเธงเธ‡เธเธฒเธฃเธ„เธฅเธฑเธ‡ เธŠเธธเธ™เธญเธดเธˆเธด เธชเธธเธ‹เธนเธเธด เน„เธ”เน‰เนเธ•เนˆเธ‡เธ•เธฑเน‰เธ‡ เน€เธ„เน‡เธ™เธˆเธด เธชเธธเธงเธฒเน‚เธ‹เน‚เธ™เธฐ เธญเธ”เธตเธ•เธญเธ˜เธดเธšเธ”เธตเธเธฃเธกเธจเธธเธฅเธเธฒเธเธฃเนเธฅเธฐเธ เธฒเธฉเธตเธชเธดเนˆเธ‡เธ™เธณเน€เธ‚เน‰เธฒเนเธซเนˆเธ‡เธเธฃเธฐเธ—เธฃเธงเธ‡เธเธฒเธฃเธ„เธฅเธฑเธ‡ เน€เธ›เน‡เธ™เธเธฃเธฃเธกเธเธฒเธฃเธšเธฃเธดเธซเธฒเธฃเธ˜เธ™เธฒเธ„เธฒเธฃเนเธซเนˆเธ‡เธ›เธฃเธฐเน€เธ—เธจเธเธตเนˆเธ›เธธเนˆเธ™เธ„เธ™เนƒเธซเธกเนˆ เธกเธตเธœเธฅเธ•เธฑเน‰เธ‡เนเธ•เนˆเธงเธฑเธ™เธ—เธตเนˆ 10 เธชเธธเธงเธฒเน‚เธ‹เน‚เธ™เธฐเธˆเธฐเธกเธฒเนเธ—เธ™ เธกเธฒเธ‹เธฒเธญเธฐเธเธด เน„เธ„เธ‹เธนเธเธฐ เธ—เธตเนˆเธžเน‰เธ™เธงเธฒเธฃเธฐเน„เธ›เนƒเธ™เธงเธฑเธ™เธ—เธตเนˆ 9 เน‚เธ”เธขเธกเธตเธงเธฒเธฃเธฐ 4 เธ›เธต

<<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:

<<References>>
2

<<Answer>>
4.4 trillion yen
Python code

from vllm import LLM, SamplingParams

llm = LLM(model="lightblue/kurage-multilingual")
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 = [
    "เธ™เธฒเธเธฒเธเธฒเธงเธฐ เธˆเธธเธ™เน‚เธเธฐ เธชเธกเธฒเธŠเธดเธเธ„เธ“เธฐเธเธฃเธฃเธกเธเธฒเธฃเธ™เน‚เธขเธšเธฒเธขเธเธฒเธฃเน€เธ‡เธดเธ™ เธ˜เธ™เธฒเธ„เธฒเธฃเนเธซเนˆเธ‡เธ›เธฃเธฐเน€เธ—เธจเธเธตเนˆเธ›เธธเนˆเธ™ เธเธฅเนˆเธฒเธงเนƒเธ™เธงเธฑเธ™เธ—เธตเนˆ 11 เธงเนˆเธฒ เธญเธฑเธ•เธฃเธฒเธ”เธญเธเน€เธšเธตเน‰เธขเธ—เธตเนˆเนเธ—เน‰เธˆเธฃเธดเธ‡เธญเธขเธนเนˆเนƒเธ™เธฃเธฐเธ”เธฑเธšเธ•เนˆเธณเธกเธฒเธเนƒเธ™เธ›เธฑเธˆเธˆเธธเธšเธฑเธ™ เนเธฅเธฐเธเธฅเนˆเธฒเธงเธงเนˆเธฒ เธซเธฒเธเนเธ™เธงเน‚เธ™เน‰เธกเน€เธจเธฃเธฉเธเธเธดเธˆเนเธฅเธฐเธฃเธฒเธ„เธฒเธ‚เธญเธ‡เธ˜เธ™เธฒเธ„เธฒเธฃเธเธฅเธฒเธ‡เธเธตเนˆเธ›เธธเนˆเธ™เน€เธ›เน‡เธ™เธˆเธฃเธดเธ‡เนƒเธ™เธญเธ™เธฒเธ„เธ• เธเธฒเธฃเธœเนˆเธญเธ™เธ„เธฅเธฒเธขเธ™เน‚เธขเธšเธฒเธขเธเธฒเธฃเน€เธ‡เธดเธ™เธˆเธฐเธ–เธนเธเธ›เธฃเธฑเธšเน‚เธ”เธขเธžเธดเธˆเธฒเธฃเธ“เธฒเธˆเธฒเธเธเธฒเธฃเธšเธฃเธฃเธฅเธธเน€เธ›เน‰เธฒเธซเธกเธฒเธขเธ”เน‰เธฒเธ™เธฃเธฒเธ„เธฒ",
    "Der Leistungsbilanzรผberschuss Japans betrug im Juli 3,2 Billionen Yen, der hรถchste monatliche รœberschuss aller Zeiten fรผr den Monat Juli. Dieser รœberschuss wird jedoch weiterhin durch das positive Primรคreinkommen unterstรผtzt, das im Juli einen รœberschuss von 4,4 Billionen Yen verzeichnete, die hรถchste monatliche Zahl in der Geschichte.",
    "้ˆดๆœจไฟŠไธ€่ฒกๅ‹™็›ธใฏ๏ผ‘๏ผๆ—ฅไป˜ใงๅ…ƒ่ฒกๅ‹™็œ้–ข็จŽๅฑ€้•ทใฎ่ซ่จชๅœ’ๅฅๅธๆฐใ‚’ๆ–ฐใ—ใ„ๆ—ฅ้Š€็†ไบ‹ใซไปปๅ‘ฝใ—ใŸใ€‚๏ผ™ๆ—ฅใซไปปๆœŸๆบ€ไบ†ใง้€€ไปปใ—ใŸ่ฒๅกšๆญฃๅฝฐๅ‰็†ไบ‹ใฎๅพŒไปปใงใ€ไปปๆœŸใฏ๏ผ”ๅนดใ€‚",
    "Lors de la phase d'apprรฉciation du yen en aoรปt, il est devenu un sujet dans le marchรฉ des changes que les investisseurs institutionnels japonais ont rรฉalisรฉ la plus grande investissement en titres ร  l'รฉtranger jamais enregistrรฉ."
]

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.

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.

To define the language of the Q&A, you can optionally include a <<Q&A Language>> control code after the text

Prompt style

Input:

<<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.

<<Q&A Language>>
English

Output:

<<Question>>
What is Japan's current account surplus in July?

<<Answer>>
3.2 trillion yen
Python code
from vllm import LLM, SamplingParams

llm = LLM(model="lightblue/kurage-multilingual")
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, language_name = None):

    str_inputs = f"""<<Q&A Generation Context>>
{context}"""

    if language_name is not None:
        str_inputs += "\n\n<<Q&A Language>>\n{language_name}"

    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

Training data

We trained on chunks sourced from the documents in 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 model.

The training data for this model can be found at lightblue/kurage_training_data.

Downloads last month
33
Safetensors
Model size
8B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for lightblue/kurage-multilingual

Quantizations
2 models

Collection including lightblue/kurage-multilingual