metadata

language:
  - en
  - ja
library_name: transformers
pipeline_tag: text-generation
license: llama2
model_type: llama

Swallow

Our Swallow model has undergone continual pre-training from the Llama 2 family, primarily with the addition of Japanese language data. The tuned versions use supervised fine-tuning (SFT). Links to other models can be found in the index.

Model Release Updates

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

April 26, 2024: Released version 0.1 of our enhanced instruction-tuned models: Swallow-7b-instruct-v0.1, Swallow-13b-instruct-v0.1, and Swallow-70b-instruct-v0.1 as preview versions.
March 2, 2024: Released the Swallow-7b-plus-hf, a model trained with approximately twice as many Japanese tokens as Swallow-7b-hf.
February 4, 2024: Released the Swallow-13b-NVE-hf.
January 26, 2024: Released the Swallow-7b-NVE-hf, Swallow-7b-NVE-instruct-hf, Swallow-70b-NVE-hf, and Swallow-70b-NVE-instruct-hf
December 19, 2024: Released the Swallow-7b-hf, Swallow-7b-instruct-hf, Swallow-13b-hf, Swallow-13b-instruct-hf, Swallow-70b-hf, and Swallow-70b-instruct-hf.

Swallow Model Index

Model	Swallow-hf	Swallow-instruct-hf	Swallow-instruct-v0.1
7B	Link	Link	Link
7B-Plus	Link	N/A	N/A
13B	Link	Link	Link
70B	Link	Link	Link

Swallow Model Index NVE (No Vocabulary Expansion)

Model	Swallow-NVE-hf	Swallow-NVE-instruct-hf
7B	Link	Link
13B	Link	N/A
70B	Link	Link

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

Model Details

Model type: Please refer to LLaMA-2 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

Instruct Model Performance

MT-Bench JA

NOTE that the models with the v0.1 suffix are newer versions compared to their original counterparts with the hf.
We will add the scores of Swallow-70b-instruct-hf and existing models soon.

Model	Average	Writing	Roleplay	Reasoning	Math	Coding	Extraction	STEM	Humanities
Swallow-7b-instruct-v0.1	0.3435	0.4450	0.4720	0.1853	0.1920	0.2204	0.3015	0.4594	0.4720
Swallow-7b-instruct-hf	0.1833	0.2205	0.1975	0.1593	0.1045	0.1282	0.2672	0.1908	0.1980
Swallow-13b-instruct-v0.1	0.3669	0.4816	0.5562	0.2769	0.1020	0.1505	0.4179	0.4347	0.5150
Swallow-13b-instruct-hf	0.2004	0.1932	0.2552	0.1507	0.1184	0.1285	0.2641	0.2434	0.2500
Swallow-70b-instruct-v0.1	0.4513	0.4822	0.5353	0.3497	0.3492	0.2668	0.5553	0.4955	0.5767
Swallow-70b-instruct-hf	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A

Evaluation Benchmarks

MT-Bench JA

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

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

Usage

First install additional dependencies in requirements.txt:

pip install -r requirements.txt

Instruction format Ver0.1

This format must be adhered to strictly, as deviations may result in less optimal outputs from the model.

The template used to construct a prompt for the Instruct model is specified as follows:

<s>[INST] <<SYS>>\n{Instruction}\n<</SYS>>\n\n{USER_MESSAGE_1} [INST] {BOT_MESSAGE_1} </s>[INST] {USER_MESSAGE_2}[/INST]

Please be aware that <s> and </s> are special tokens used for the beginning of string (BOS) and end of string (EOS), respectively, while [INST] and [/INST] are considered regular strings.

For the "{Instruction}" part, We recommend using "あなたは誠実で優秀な日本人のアシスタントです。"

Use the instruct model Ver0.1

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "tokyotech-llm/Swallow-70b-instruct-v0.1"
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_name)

device = "cuda"

messages = [
    {"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": "東京工業大学の主なキャンパスについて教えてください"}
]

encodeds = tokenizer.apply_chat_template(messages, return_tensors="pt")

model_inputs = encodeds.to(device)
model.to(device)

generated_ids = model.generate(model_inputs, max_new_tokens=128, do_sample=True)
decoded = tokenizer.batch_decode(generated_ids)
print(decoded[0])

Training Datasets

Instruction Tuning Ver0.1

The following datasets were used for the instruction tuning.

OpenAssistant Conversations Dataset EN top-1 thread
OpenAssistant Conversations Dataset was used, where human utterances are included but the responses are not used. Instead, the responses were generated using the Mixtral-8x7B-Instruct-v0.1 model.

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 2 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

Authors

Here are the team members:

From Okazaki Laboratory, the following members:
From YOKOTA Laboratory, the following members: