Llama 3.1 Swallow - Built with Llama

Llama 3.1 Swallow is a series of large language models (8B, 70B) that were built by continual pre-training on the Meta Llama 3.1 models. Llama 3.1 Swallow enhanced the Japanese language capabilities of the original Llama 3.1 while retaining the English language capabilities. We use approximately 200 billion tokens that were sampled from a large Japanese web corpus (Swallow Corpus Version 2), Japanese and English Wikipedia articles, and mathematical and coding contents, etc (see the Training Datasets section of the base model) for continual pre-training. The instruction-tuned models (Instruct) were built by supervised fine-tuning (SFT) on the synthetic data specially built for Japanese. See the Swallow Model Index section to find other model variants.

Note: Llama-3.1-Swallow-8B-Instruct-v0.3 model was continually pre-trained from the meta-llama/Llama-3.1-8B-Instruct and then instruction-tuned with our instruction datasets.

Release History

• December 23, 2024: Released Llama-3.1-Swallow-8B-Instruct-v0.3.
• November 11, 2024: Released Llama-3.1-Swallow-8B-v0.2 and Llama-3.1-Swallow-8B-Instruct-v0.2.
• October 08, 2024: Released Llama-3.1-Swallow-8B-v0.1, Llama-3.1-Swallow-8B-Instruct-v0.1, Llama-3.1-Swallow-70B-v0.1, and Llama-3.1-Swallow-70B-Instruct-v0.1.

Major Updates

This release enhances the conversation capability of Llama 3.1 Swallow. The updated model, Llama-3.1-Swallow-8B-Instruct-v0.3 generates helpful and detailed responses based on user instructions and conversation history. Among all open-source LLMs with <= 8 billion parameters, Llama-3.1-Swallow-8B-Instruct-v0.3 exhibits state-of-the-art performance on Japanese MT-Bench, outperforming its predecessor, Llama-3.1-Swallow-8B-Instruct-v0.2, by 8.4 points.

Swallow Model Index

The website https://swallow-llm.github.io/ provides large language models developed by the Swallow team.

Model Details

• Model type: Please refer to Llama 3.1 MODEL_CARD for details on the model architecture.
• Language(s): Japanese English
• Library: Megatron-LM
• Tokenizer: Please refer to Llama 3.1 blog for details on the tokenizer.
• Contact: swallow[at]nlp.c.titech.ac.jp

Model Performance

MT-Bench JA

Japanese tasks

English tasks

Evaluation Benchmarks

MT-Bench JA

We used Japanese MT-Bench to assess the capabilities of multi-turn dialogue with the following settings:

• Implementation: FastChat [Zheng+, 2023] (commit #e86e70d0)
• Question: Nejumi LLM-Leaderboard NEO, mtbench_ja_question_v3
• Reference Answer: Nejumi LLM-Leaderboard NEO, mtbench_ja_referenceanswer_v1
• Prompt for Judge: Nejumi LLM-Leaderboard NEO, mtbench_ja_prompt_v1
• Judge: gpt-4-1106-preview
• Scoring: Absolute scale normalized to a 0-1 range, averaged over five runs.

Japanese evaluation benchmarks

We used llm-jp-eval(v1.3.0), JP Language Model Evaluation Harness(commit #9b42d41) and Code Generation LM Evaluation Harness(commit #0261c52). The details are as follows:

• Multiple-choice question answering (JCommonsenseQA [Kurihara et al., 2022])
• Open-ended question answering (JEMHopQA [Ishii et al., 2024])
• Open-ended question answering (NIILC [関根, 2003])
• Machine reading comprehension (JSQuAD [Kurihara et al., 2022])
• Automatic summarization (XL-Sum [Hasan et al., 2021])
• Machine translation (WMT2020 ja-en [Barrault et al., 2020])
• Machine translation (WMT2020 en-ja [Barrault et al., 2020])
• Mathematical reasoning (MGSM [Shi et al., 2023])
• Academic exams (JMMLU [尹ら, 2024])
• Code generation (JHumanEval [佐藤ら, 2024])

English evaluation benchmarks

We used the Language Model Evaluation Harness(v.0.4.2) and Code Generation LM Evaluation Harness(commit #0261c52). The details are as follows:

• Multiple-choice question answering (OpenBookQA [Mihaylov et al., 2018])
• Open-ended question answering (TriviaQA [Joshi et al., 2017])
• Machine reading comprehension (SQuAD2 [Rajpurkar et al., 2018])
• Commonsense reasoning (XWINO [Tikhonov and Ryabinin, 2021])
• Natural language inference (HellaSwag [Zellers et al., 2019])
• Mathematical reasoning (GSM8K [Cobbe et al., 2021])
• Reasoning (BBH (BIG-Bench-Hard) [Suzgun et al., 2023])
• Academic exams (MMLU [Hendrycks et al., 2021])
• Code generation (HumanEval [Chen et al., 2021])

Usage

pip install vllm

from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

model_name = "tokyotech-llm/Llama-3.1-Swallow-8B-Instruct-v0.3"

tokenizer = AutoTokenizer.from_pretrained(model_name)
llm = LLM(
    model=model_name,
    tensor_parallel_size=1,
)

sampling_params = SamplingParams(
    temperature=0.6, top_p=0.9, max_tokens=512, stop="<|eot_id|>"
)


message = [
    {"role": "system", "content": "あなたは誠実で優秀な日本人のアシスタントです。"},
    {
        "role": "user",
        "content": "東京の紅葉した公園で、東京タワーと高層ビルを背景に、空を舞うツバメと草地に佇むラマが出会う温かな物語を書いてください。",
    },
]
prompt = tokenizer.apply_chat_template(
    message, tokenize=False, add_generation_prompt=True
)

output = llm.generate(prompt, sampling_params)

print(output[0].outputs[0].text)

Training Datasets

Instruction Tuning

The following datasets were used for the instruction tuning.

• Gemma-2-LMSYS-Chat-1M-Synth
  • Multi-turn Japanese instruction dataset synthesized and derived from lmsys-chat-1m [Zhang+, ICLR24]).
  • First-turn user instructions were translated into Japanese via DeepL (machine translation), and assistant responses were generated using gemma-2-27b-it. The same model, i.e., gemma-2-27b-it served as a judge for rejection sampling (n=6).
  • Second-turn user instructions and responses were synthesized using gemma-2-27b-it. The same model scores the quality of the second-turn response with a range of 1-10. Second-turn responses with scores lower than 9 were rejected, along with their corresponding instructions.
    Conversations containing personally identifiable information (PII) and template-based user instructions were removed. Duplicate instructions were removed.
• Swallow-Magpie-Ultra-v0.1
  • A Japanese variant of the filtered-magpie-ultra-en dataset, translated into Japanese by gemma-2-27b-it.
• Swallow-Gemma-Magpie-v0.1
  • A Japanese synthetic instruction tuning dataset from scratch, generated by gemma-2-27b-it. User instructions were created with prompts specific to each topic, and assistant responses were generated for these instructions.
  • The conversations were heuristically filtered for quality and length. Then, gemma-2-27b-it was applied to score the quality of each of the conversation with a range of 1-10. Conversations with scores <= 7 were rejected.

Risks and Limitations

The models released here are still in the early stages of our research and development and have not been tuned to ensure outputs align with human intent and safety considerations.

Acknowledgements

We thank Meta Research for releasing Llama 3.1 under a generous open license.

We received various supports, including:

• AIST project: "Research and Development of Foundation Models for Generative AI in the Physical Domain"
• NEDO project: "Development of Artificial Intelligence Application Technology to Support Judgment in Design Risk Assessment Work Based on the Perspective of Skilled Persons" (JPNP18002) of "Development of Integration Technology as the Core of Next Generation Artificial Intelligence and Robotics"
• MEXT project: "Formation of R&D center to ensure transparency and reliability of generative AI models"
• AIST program: Large Generative AI Development Support Program

License

META LLAMA 3.1 COMMUNITY LICENSE and Gemma Terms of Use

Authors

Here are the team members:

• From Tokyo Institute of Technology Okazaki Laboratory, the following members:
  • Naoaki Okazaki
  • Sakae Mizuki
  • Youmi Ma
  • Koki Maeda
  • Kakeru Hattori
  • Masanari Ohi
  • Hinari Shimada
  • Taihei Shiotani
  • Koshiro Saito
• From Tokyo Institute of Technology YOKOTA Laboratory, the following members:
  • Rio Yokota
  • Kazuki Fujii
  • Taishi Nakamura
  • Takumi Okamoto
  • Ishida Shigeki
• From Artificial Intelligence Research Center, AIST, Japan, the following members:
  • Hiroya Takamura

How to cite

If you find our work helpful, please feel free to cite these papers.

@inproceedings{Fujii:COLM2024,
   title={Continual Pre-Training for Cross-Lingual LLM Adaptation:
Enhancing Japanese Language Capabilities},
   author={Kazuki Fujii and Taishi Nakamura and Mengsay Loem and Hiroki
Iida and Masanari Ohi and Kakeru Hattori and Hirai Shota and Sakae
Mizuki and Rio Yokota and Naoaki Okazaki},
   booktitle="Proceedings of the First Conference on Language Modeling",
   series={COLM},
   pages="(to appear)",
   year="2024",
   month=oct,
   address={University of Pennsylvania, USA},
}

@inproceedings{Okazaki:COLM2024,
   title={Building a Large Japanese Web Corpus for Large Language Models},
   author={Naoaki Okazaki and Kakeru Hattori and Hirai Shota and Hiroki
Iida and Masanari Ohi and Kazuki Fujii and Taishi Nakamura and Mengsay
Loem and Rio Yokota and Sakae Mizuki},
   booktitle="Proceedings of the First Conference on Language Modeling",
   series={COLM},
   pages="(to appear)",
   year="2024",
   month=oct,
   address={University of Pennsylvania, USA},
}

@misc{ma:arxiv2025,
      title={Building Instruction-Tuning Datasets from Human-Written Instructions with Open-Weight Large Language Models}, 
      author={Youmi Ma and Sakae Mizuki and Kazuki Fujii and Taishi Nakamura and Masanari Ohi and Hinari Shimada and Taihei Shiotani and Koshiro Saito and Koki Maeda and Kakeru Hattori and Takumi Okamoto and Shigeki Ishida and Rio Yokota and Hiroya Takamura and Naoaki Okazaki},
      year={2025},
      eprint={2503.23714},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2503.23714}, 
}

References

@misc{dubey2024llama3herdmodels,
      title={The Llama 3 Herd of Models}, 
      author={Abhimanyu Dubey and Abhinav Jauhri and Abhinav Pandey and Abhishek Kadian and Ahmad Al-Dahle and Aiesha Letman and Akhil Mathur and Alan Schelten and Amy Yang and Angela Fan et al.},
      year={2024},
      eprint={2407.21783},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2407.21783}, 
}