Swallow-MS-7b-v0.1

Our Swallow-MS-7b-v0.1 model has undergone continual pre-training from the Mistral-7B-v0.1, primarily with the addition of Japanese language data.

Model Release Updates

We are excited to share the release schedule for our latest models:

• April 26, 2024: Released the Swallow-MS-7b-instruct-v0.1
• March 11, 2024: Released the Swallow-MS-7b-v0.1

This repository provides large language models developed by TokyoTech-LLM.

Model Details

• Model type: Please refer to Mistral technical report for details on the model architecture.
• Language(s): Japanese English
• Tokenizer: This model employs a tokenizer that features a broadened vocabulary based on Japanese data. This allows for a more efficient representation of text using fewer tokens, leading to a notably faster inference process.
• Contact: swallow[at]nlp.c.titech.ac.jp

Base Model Performance

Japanese tasks

Model	Size	JCommonsenseQA	JEMHopQA	NIILC	JSQuAD	XL-Sum	MGSM	WMT20-en-ja	WMT20-ja-en	Average
		4-shot	4-shot	4-shot	4-shot	1-shot	4-shot	4-shot	4-shot
CyberAgentLM2-7B	7B	0.2198	0.5047	0.5066	0.7799	0.0233	0.0600	0.2345	0.1499	0.3098
Llama 2	7B	0.3852	0.4240	0.3410	0.7917	0.1905	0.0760	0.1783	0.1738	0.3201
japanese-stablelm-base-beta-7b	7B	0.3610	0.4478	0.4432	0.8318	0.2195	0.0720	0.1946	0.1226	0.3366
japanese-stablelm-base-ja_vocab-beta-7b	7B	0.2172	0.4482	0.4309	0.8202	0.0757	0.0520	0.1601	0.1453	0.2937
ELYZA-japanese-Llama-2-7b	7B	0.5791	0.4703	0.4019	0.8226	0.1312	0.0600	0.1795	0.1289	0.3467
ELYZA-japanese-Llama-2-7b-fast	7B	0.5308	0.4330	0.3898	0.8131	0.1289	0.0720	0.1678	0.1143	0.3312
youri-7b (base)	7B	0.4620	0.4776	0.4999	0.8506	0.1957	0.0640	0.2671	0.1971	0.3768
Swallow-7b	7B	0.4808	0.5078	0.5968	0.8573	0.1830	0.1240	0.2510	0.1511	0.3940
Swallow-7b-plus	7B	0.5478	0.5493	0.6030	0.8544	0.1806	0.1360	0.2568	0.1441	0.4090
Qwen-7B	7B	0.7712	0.4234	0.2376	0.8594	0.1371	0.2160	0.1689	0.1801	0.3742
nekomata-7b	7B	0.7417	0.4928	0.5022	0.8707	0.1676	0.1240	0.2673	0.1815	0.4185
Mistral-7B-v0.1	7B	0.7301	0.4245	0.2722	0.8563	0.2006	0.1760	0.1405	0.1733	0.3717
japanese-stablelm-base-gamma-7b	7B	0.7364	0.4643	0.5568	0.8910	0.2293	0.1680	0.2390	0.1561	0.4301
Swallow-MS-7b-v0.1	7B	0.8570	0.4915	0.5519	0.8802	0.1988	0.2240	0.2494	0.1667	0.4524

English tasks

Model	Size	OpenBookQA	TriviaQA	HellaSwag	SQuAD2.0	XWINO	GSM8K	Average
		8-shot	8-shot	8-shot	8-shot	8-shot	8-shot
CyberAgentLM2-7B	7B	0.2860	0.3496	0.5003	0.3510	0.8581	0.0705	0.4026
Llama 2	7B	0.3580	0.6265	0.5860	0.3207	0.9049	0.1410	0.4895
japanese-stablelm-base-beta-7b	7B	0.3620	0.5903	0.5707	0.2992	0.8994	0.1198	0.4736
japanese-stablelm-base-ja_vocab-beta-7b	7B	0.3520	0.5549	0.5644	0.3079	0.8942	0.0538	0.4545
ELYZA-japanese-Llama-2-7b	7B	0.3400	0.5875	0.5595	0.2721	0.8989	0.1638	0.4703
ELYZA-japanese-Llama-2-7b-fast	7B	0.3280	0.5817	0.5530	0.2605	0.8989	0.1425	0.4608
youri-7b (base)	7B	0.3400	0.5257	0.5540	0.3297	0.8938	0.0963	0.4566
Swallow-7b	7B	0.3180	0.4836	0.5308	0.3125	0.8817	0.1130	0.4399
Swallow-7b-plus	7B	0.3280	0.4558	0.5259	0.3134	0.8929	0.1061	0.4370
Qwen-7B	7B	0.3640	0.5695	0.5787	0.3799	0.8933	0.4617	0.5412
nekomata-7b	7B	0.3340	0.4371	0.5340	0.2933	0.8766	0.1531	0.4380
Mistral-7B-v0.1	7B	0.3660	0.7050	0.6264	0.3799	0.9157	0.3533	0.5577
japanese-stablelm-base-gamma-7b	7B	0.3240	0.5745	0.5739	0.3546	0.8976	0.1911	0.4860
Swallow-MS-7b-v0.1	7B	0.3440	0.5976	0.5810	0.3364	0.9037	0.2623	0.5042

Code generation tasks

Model	Size	JHumanEval	HumanEval
		pass@1	pass@1
CyberAgentLM2-7B	7B	0.0634	0.0756
Llama 2	7B	0.1152	0.1378
japanese-stablelm-base-beta-7b	7B	0.1018	0.1280
japanese-stablelm-base-ja_vocab-beta-7b	7B	0.0896	0.1122
ELYZA-japanese-Llama-2-7b	7B	0.0287	0.0427
ELYZA-japanese-Llama-2-7b-fast	7B	0.0000	0.0037
youri-7b (base)	7B	0.0829	0.0982
Swallow-7b	7B	0.0183	0.0183
Swallow-7b-plus	7B	0.0061	0.0037
Qwen-7B	7B	0.1701	0.1805
nekomata-7b	7B	0.0988	0.1402
Mistral-7B-v0.1	7B	0.2555	0.2933
japanese-stablelm-base-gamma-7b	7B	0.1823	0.1915
Swallow-MS-7b-v0.1	7B	0.2305	0.2768

Evaluation Benchmarks

Japanese evaluation benchmarks

We used llm-jp-eval(v1.0.0) and JP Language Model Evaluation Harness(commit #9b42d41). The details are as follows:

• Multiple-choice question answering (JCommonsenseQA [Kurihara+, 2022])
• Open-ended question answering (JEMHopQA [Ishii+, 2023])
• Open-ended question answering (NIILC [Sekine, 2003])
• Machine reading comprehension (JSQuAD [Kurihara+, 2022])
• Automatic summarization (XL-Sum [Hasan+, 2021])
• Machine translation (WMT2020 ja-en [Barrault+, 2020])
• Machine translation (WMT2020 en-ja [Barrault+, 2020])
• Mathematical reasoning (MGSM [Shi+, 2023])

English evaluation benchmarks

We used the Language Model Evaluation Harness(v.0.3.0). The details are as follows:

• Multiple-choice question answering (OpenBookQA [Mihaylov+, 2018])
• Open-ended question answering (TriviaQA [Joshi+, 2017])
• Machine reading comprehension (SQuAD 2.0 [Rajpurkar+, 2018])
• Commonsense reasoning (XWINO [Tikhonov & Ryabinin, 2021])
• Natural language inference (HellaSwag [Zellers+, 2019])
• Mathematical reasoning (GSM8k [Cobbe+, 2021])

Code evaluation benchmarks

We utilized the Code Generation LM Evaluation Harness [Allal+, 2022] (commit #0261c52). The details are as follows:

• Code generation (HumanEval [Chen+, 2021])
• Code generation in Japanese (JHumanEval [Satoh+, 2024])

Usage

First install additional dependencies in requirements.txt:

pip install -r requirements.txt

Use the base model

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "tokyotech-llm/Swallow-MS-7b-v0.1"
tokenizer = AutoTokenizer.from_pretrained(model_name)

model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16, device_map="auto")
prompt = "東京工業大学の主なキャンパスは、"
input_ids = tokenizer.encode(
    prompt,
    add_special_tokens=False,
    return_tensors="pt"
)
tokens = model.generate(
    input_ids.to(device=model.device),
    max_new_tokens=128,
    temperature=0.99,
    top_p=0.95,
    do_sample=True,
)

out = tokenizer.decode(tokens[0], skip_special_tokens=True)
print(out)

Training Datasets

Continual Pre-Training

The following datasets were used for continual pre-training.

• Algebraic Stack
• Japanese Wikipedia
• RefinedWeb
• Swallow Corpus
• The Pile

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 Mistral AI for releasing Mistral 7B v0.1 under an open license for others to build on.

Our project is supported by the ABCI Large-scale Language Model Building Support Program of the National Institute of Advanced Industrial Science and Technology.

License

apache-2.0

Authors

Here are the team members:

• From Okazaki Laboratory, the following members:
  • Naoaki Okazaki
  • Sakae Mizuki
  • Hiroki Iida
  • Mengsay Loem
  • Shota Hirai
  • Kakeru Hattori
  • Masanari Ohi
• From YOKOTA Laboratory, the following members:
  • Rio Yokota
  • Kazuki Fujii
  • Taishi Nakamura

How to cite

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

@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},
}