Llama 3.1 Swallow

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) 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.

Release History

• 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.

Swallow Model Index

Model	Llama-3.1-Swallow	Llama-3.1-Swallow-Instruct
8B	Link	Link
70B	Link	Link

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

Third-Party Community Consideration

This model is not owned or developed by NVIDIA. This model has been developed and built to a third-party’s requirements for this application and use case; see link to Non-NVIDIA Model Card.

License and Terms of Use

GOVERNING TERMS: This trial service is governed by the NVIDIA API Trial Terms of Service. Use of this model is governed by the Llama 3.1 Community License Agreement and the Gemma Terms of Use. Built with Llama.

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 Architecture:

Architecture Type: Transformer

Input:

Input Type(s): Text
Input Format(s): String
Input Parameters: One Dimensional (1D)

Output:

Output Type(s): Text
Output Format: String
Output Parameters: 1D

Supported Hardware Microarchitecture Compatibility:

• NVIDIA Ampere
  
• NVIDIA Hopper
  
• NVIDIA Lovelace
  

Model Performance

Japanese tasks

Model	JCom.	JEMHopQA	NIILC	JSQuAD	XL-Sum	MGSM	WMT20-en-ja	WMT20-ja-en	JMMLU	JHumanEval	Ja Avg
	4-shot	4-shot	4-shot	4-shot	1-shot	4-shot	4-shot	4-shot	5-shot	0-shot
	EM acc	Char-F1	Char-F1	Char-F1	ROUGE-2	EM acc	BLEU	BLEU	EM acc	pass@1
Qwen2-72B-Instruct	0.9634	0.6268	0.5418	0.9210	0.1644	0.7840	0.2592	0.2327	0.7713	0.6909	0.5955
Qwen2.5-72B-Instruct	0.9696	0.5699	0.5811	0.7381	0.1706	0.8360	0.2269	0.2179	0.7899	0.6256	0.5726
Llama 3 70B Instruct	0.9419	0.6114	0.5506	0.9164	0.1912	0.7200	0.2708	0.2350	0.6789	0.6610	0.5777
Llama 3.1 70B Instruct	0.9482	0.6246	0.5781	0.9201	0.1772	0.7440	0.2805	0.2472	0.7323	0.6933	0.5945
Llama 3 Youko 70B Instruct	0.9526	0.6252	0.5853	0.9215	0.1983	0.7400	0.2633	0.2245	0.7170	0.6098	0.5838
Llama-3.1-70B-Japanese-Instruct-2407	0.9562	0.6466	0.6602	0.9187	0.1564	0.7480	0.2901	0.2410	0.7227	0.6274	0.5967
Llama 3 heron brain 70B v0.3	0.9660	0.6643	0.6817	0.9221	0.2611	0.7720	0.3093	0.2578	0.7077	0.6079	0.6150
Llama 3 Swallow 70B Instruct	0.9607	0.6188	0.6026	0.9236	0.1389	0.6560	0.2724	0.2532	0.6572	0.6000	0.5683
Llama 3.1 Swallow 70B Instruct	0.9598	0.6192	0.6605	0.9235	0.1938	0.7760	0.3123	0.2593	0.7117	0.4713	0.5887

English tasks

Model	OpenBookQA	TriviaQA	HellaSWAG	SQuAD2.0	XWINO	MMLU	GSM8K	BBH	HumanEval	En Avg
	4-shot	4-shot	4-shot	4-shot	4-shot	5-shot	4-shot	3-shot	0-shot
	Acc	EM acc	Acc	EM acc	Acc	Acc	EM acc	CoT EM Acc	pass@1
Qwen2-72B-Instruct	0.4360	0.7588	0.6857	0.3913	0.9110	0.8391	0.8499	0.2436	0.6939	0.6455
Qwen2.5-72B-Instruct	0.4540	0.6764	0.7064	0.3550	0.8895	0.8478	0.9113	0.4027	0.6165	0.6511
Llama 3 70B Instruct	0.4400	0.7999	0.6552	0.4024	0.9127	0.7992	0.9052	0.8326	0.7555	0.7225
Llama 3.1 70B Instruct	0.4300	0.8212	0.6621	0.3921	0.9157	0.8213	0.8764	0.8390	0.7915	0.7277
Llama 3 Youko 70B Instruct	0.4500	0.7973	0.6863	0.3914	0.9153	0.8055	0.8923	0.7814	0.6598	0.7088
Llama-3.1-70B-Japanese-Instruct-2407	0.4220	0.8104	0.6481	0.3744	0.9170	0.8071	0.8893	0.8228	0.7463	0.7153
Llama 3 heron brain 70B v0.3	0.4460	0.8107	0.6682	0.4085	0.9174	0.7898	0.8772	0.7586	0.6713	0.7053
Llama 3 Swallow 70B Instruct	0.4520	0.8174	0.6758	0.4050	0.9230	0.7883	0.8688	0.8152	0.6890	0.7150
Llama 3.1 Swallow 70B Instruct	0.4520	0.8148	0.6834	0.4012	0.9157	0.7855	0.8886	0.8486	0.5823	0.7080

MT-Bench JA

Model	coding	extraction	humanities	math	reasoning	roleplay	stem	writing	JMTAvg
Qwen2-72B-Instruct	0.5699	0.7858	0.8222	0.5096	0.7032	0.7963	0.7728	0.8223	0.7228
Qwen2.5-72B-Instruct	0.7060	0.7866	0.8122	0.6968	0.6536	0.8301	0.8060	0.7841	0.7594
Llama 3 70B Instruct	0.5969	0.8410	0.7120	0.4481	0.4884	0.7117	0.6510	0.6900	0.6424
Llama 3.1 70B Instruct	0.5252	0.7846	0.7086	0.5063	0.6979	0.6888	0.6402	0.6653	0.6521
Llama 3 Youko 70B Instruct	0.6632	0.8387	0.8108	0.4655	0.7013	0.7778	0.7544	0.7662	0.7222
Llama-3.1-70B-Japanese-Instruct-2407	0.6267	0.7525	0.7938	0.5750	0.5590	0.7725	0.7240	0.7180	0.6902
Llama 3 heron brain 70B v0.3	0.3762	0.7892	0.7274	0.5589	0.5070	0.6662	0.6880	0.6996	0.6266
Llama 3 Swallow 70B Instruct	0.5269	0.7250	0.5690	0.4669	0.6121	0.6238	0.5533	0.5698	0.5809
Llama 3.1 Swallow 70B Instruct	0.5676	0.7859	0.7490	0.5437	0.6383	0.6870	0.6121	0.6540	0.6547
GPT-3.5 (gpt-3.5-turbo-0125)	0.6851	0.7641	0.7414	0.5522	0.5128	0.7104	0.6266	0.7361	0.6661
GPT-4o (gpt-4o-2024-05-13)	0.7296	0.8540	0.8646	0.6641	0.6661	0.8274	0.8184	0.8085	0.7791

Evaluation Benchmarks

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])

MT-Bench JA

We used Japanese MT-Bench to assess the instruction-following capabilities of models. We utilized the following settings:

• Implemantation: 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-Lederboard NEO, mtbench_ja_prompt_v1
• Judge: gpt-4-1106-preview
• Scoring: Absolute scale normalized to a 0-1 range, averaged over five runs.

Inference:

Engine: TensorRT-LLM
Test Hardware:

• NVIDIA H100x4
  

Ethical Considerations:

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please report security vulnerabilities or NVIDIA AI Concerns here.

Usage

Copy Code
pip install vllm

Copy Code
from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

model_name = "tokyotech-llm/Llama-3.1-Swallow-70B-Instruct-v0.1"

tokenizer = AutoTokenizer.from_pretrained(model_name)
llm = LLM(
    model=model_name,
    tensor_parallel_size=4,
)

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 instruction datasets were used for the instruction tuning.

• Japanese
  • lmsys-chat-1m-synth-ja-wo-pii-and-template-instructions
    • Single-turn Japanese synthetic instruction dataset derived from lmsys-chat-1m dataset [Zhang+, ICLR24]). The first-turn user instructions were translated into Japanese via DeepL machine translation, and the assistant responses were generated using the Llama-3.1-405B-Instruct model. Rejection sampling (n=6) was applied, with Llama-3.1-70B-Instruct serving as a judge.
    • As implied by the dataset name, conversations that contain personally identifiable information (PII) or template-based user instructions have been removed. Duplicate instuctions have also been removed.
  • filtered-magpie-ultra-ja
    • A Japanese variant of the filtered-magpie-ultra-en dataset, machine-translated into Japanese using the gemma-2-27b-it.
  • gemma-magpie
    • A Japanese synthetic Q&A dataset from scratch, generated using gemma-2-27b-it. User instructions were created with prompts specific to each topic, and the assistant responses were generated for these instructions. The conversations were then heuristically filtered for quality and length.
• English
  • lmsys-chat-1m-synth-en-wo-pii-and-template-instructions
    • Similar to the lmsys-chat-1m-synth-ja-wo-pii-and-template-instructions, but this version uses the original English user instructions. The assistant responses were generated in English as well. Rejection sampling was not applied in this version.
  • filtered-magpie-ultra-en
    • A subset of the magpie-ultra dataset, developed following the MAGPIE recipe [Xu+, arXiv24] using Llama-3.1-405B-Instruct. This subset includes only samples rated as 'average,' 'good,' or 'excellent.'

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

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
  • 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 us.

Copy Code
@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},
}

References

Copy Code
@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}, 
}