metadata

license: apache-2.0
language:
  - en
  - ja
programming_language:
  - C
  - C++
  - C#
  - Go
  - Java
  - JavaScript
  - Lua
  - PHP
  - Python
  - Ruby
  - Rust
  - Scala
  - TypeScript
pipeline_tag: text-generation
library_name: transformers
inference: false

llm-jp-3-8x1.8b-instruct3

LLM-jp-3 is the series of large language models developed by the Research and Development Center for Large Language Models at the National Institute of Informatics.

This repository provides the llm-jp-3-8x1.8b-instruct3 model. For an overview of the LLM-jp-3 models across different parameter sizes, please refer to:

Checkpoints format: Hugging Face Transformers

Required Libraries and Their Versions

torch>=2.3.0
transformers>=4.40.1
tokenizers>=0.19.1
accelerate>=0.29.3
flash-attn>=2.5.8

Usage

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("llm-jp/llm-jp-3-8x1.8b-instruct3")
model = AutoModelForCausalLM.from_pretrained("llm-jp/llm-jp-3-8x1.8b-instruct3", device_map="auto", torch_dtype=torch.bfloat16)
chat = [
    {"role": "system", "content": "以下は、タスクを説明する指示です。要求を適切に満たす応答を書きなさい。"},
    {"role": "user", "content": "自然言語処理とは何か"},
]
tokenized_input = tokenizer.apply_chat_template(chat, add_generation_prompt=True, tokenize=True, return_tensors="pt").to(model.device)
with torch.no_grad():
    output = model.generate(
        tokenized_input,
        max_new_tokens=100,
        do_sample=True,
        top_p=0.95,
        temperature=0.7,
        repetition_penalty=1.05,
    )[0]
print(tokenizer.decode(output))

Model Details

Model type: Transformer-based Language Model
Total seen tokens: 2.1T tokens

Params	Layers	Hidden size	Heads	Routed Experts	Activated Experts	Context length	Embedding parameters	Non-embedding parameters	Activated parameters	Total parameters
150M	12	512	8	-	-	4096	101,874,688	50,344,448	152,219,136	152,219,136
440M	16	1024	8	-	-	4096	203,749,376	243,303,424	447,052,800	447,052,800
980M	20	1536	8	-	-	4096	305,624,064	684,258,816	989,882,880	989,882,880
1.8b	24	2048	16	-	-	4096	407,498,752	1,459,718,144	1,867,216,896	1,867,216,896
8x1.8b	24	2048	16	8	2	4096	407,498,752	8,858,863,616	2,924,279,808	9,266,362,368
3.7b	28	3072	24	-	-	4096	611,248,128	3,171,068,928	3,782,317,056	3,782,317,056
7.2b	32	4096	32	-	-	4096	814,997,504	6,476,271,616	7,291,269,120	7,291,269,120
13b	40	5120	40	-	-	4096	1,018,746,880	12,688,184,320	13,706,931,200	13,706,931,200
8x13b	40	5120	40	8	2	4096	1,018,746,880	72,144,081,920	22,200,806,400	73,162,828,800
172b	96	12288	96	-	-	4096	2,444,992,512	169,947,181,056	172,392,173,568	172,392,173,568

Tokenizer

The tokenizer of this model is based on huggingface/tokenizers Unigram byte-fallback model. The vocabulary entries were converted from llm-jp-tokenizer v3.0. Please refer to README.md of llm-jp-tokenizer for details on the vocabulary construction procedure (the pure SentencePiece training does not reproduce our vocabulary).

Datasets

Pre-training

The models have been pre-trained using a blend of the following datasets.

Language	Dataset	Tokens
Japanese	Wikipedia	2.6B
	Common Crawl	762.8B
	WARP/PDF	237.3B
	WARP/HTML	2.7B
	Kaken	1.8B
English	Wikipedia	4.7B
	Dolma/CC-head	608.5B
	Dolma/C4	181.6B
	Dolma/Reddit	83.1B
	Dolma/PeS2o	62.9B
	Dolma/Gutenberg	5.5B
	Dolma/Wiki	3.9B
Code	The Stack	114.1B
Chinese	Wikipedia	0.8B
Korean	Wikipedia	0.3B

Post-training

We have fine-tuned the pre-trained checkpoint with supervised fine-tuning and further aligned it with Direct Preference Optimization.

Supervised Fine-tuning

The datasets used for supervised fine-tuning are as follows:

Language	Dataset	Description
Japanese	ichikara-instruction-004-002	A manually constructed instruction dataset.
	AnswerCarefully (ver2.0)	A manually constructed instruction dataset focusing on LLMs' safety.
	ichikara-instruction-format	A small subset of the ichikara-instruction dataset, edited with some constraints on the output format.
	AutoMultiTurnByCalm3-22B	A synthetic instruction dataset.
	ramdom-to-fixed-multiturn-Calm3	A synthetic instruction dataset.
	wizardlm8x22b-logical-math-coding-sft-ja	A synthetic instruction dataset.
	magpie-sft-v1.0	A synthetic instruction dataset we created.
English	Daring-Anteater	-
	FLAN	-
Japanese & English	Synthetic-JP-EN-Coding-Dataset	A synthetic instruction dataset.

Direct Preference Optimization

The datasets used for supervised fine-tuning are as follows:

Language	Dataset	Description
Japanese	aya-ja-evol-inst	A synthetic preference dataset focusing on LLMs' helpfulness.
	ac-self-inst	A synthetic preference dataset focusing on LLMs' safety.

Evaluation

llm-jp-eval (v1.4.1)

We evaluated the models using 100 examples from the dev split. Note that we skipped the CG (Code Generation) task.

Model name	average	EL	FA	HE	MC	MR	MT	NLI	QA	RC	SUM
llm-jp/llm-jp-3-8x1.8b	0.4538	0.3868	0.2412	0.265	0.5300	0.5100	0.8097	0.4760	0.5373	0.7553	0.0262
llm-jp/llm-jp-3-8x1.8b-instruct2	0.5130	0.4475	0.2296	0.4050	0.6433	0.5600	0.8150	0.5660	0.5611	0.8366	0.0660
llm-jp/llm-jp-3-8x1.8b-instruct3	0.5145	0.4515	0.2265	0.4250	0.6833	0.5400	0.8212	0.5580	0.5453	0.8193	0.0753
llm-jp/llm-jp-3-8x13b	0.5867	0.5445	0.2913	0.4950	0.8033	0.7200	0.8378	0.5780	0.6461	0.8543	0.0966
llm-jp/llm-jp-3-8x13b-instruct2	0.6256	0.5521	0.2889	0.5250	0.8967	0.7500	0.8362	0.6820	0.6367	0.9065	0.1816
llm-jp/llm-jp-3-8x13b-instruct3	0.6246	0.5477	0.2852	0.5250	0.9067	0.7600	0.8386	0.6880	0.6267	0.9041	0.1635

Japanese MT Bench

We evaluated the models using gpt-4o-2024-08-06. The scores represent the average values obtained from five rounds of inference and evaluation. For more details, please refer to the codes.

Model name	average	coding	extraction	humanities	math	reasoning	roleplay	stem	writing
llm-jp/llm-jp-3-172b-instruct3	6.36	4.24	6.66	8.11	4.58	5.74	7.44	6.76	7.36
llm-jp/llm-jp-3-8x1.8b-instruct2	5.47	3.47	4.90	7.78	3.51	4.38	6.84	6.35	6.54
llm-jp/llm-jp-3-8x1.8b-instruct3	5.52	3.60	5.23	7.81	3.87	4.53	6.40	5.98	6.72
llm-jp/llm-jp-3-8x13b-instruct2	6.62	4.50	6.53	8.56	5.30	6.03	7.86	7.10	7.12
llm-jp/llm-jp-3-8x13b-instruct3	6.58	4.90	6.41	8.32	5.37	5.20	7.75	7.24	7.48

AnswerCarefully-Eval

AnswerCarefully-Eval assesses the safety of Japanese language model outputs using the LLM-as-a-Judge approach, based on the test set from llm-jp/AnswerCarefully. We evaluated the models using gpt-4-0613. The scores represent the average values obtained from five rounds of inference and evaluation.

Model name	Acceptance rate (%, ↑)	Violation rate (%, ↓)
llm-jp/llm-jp-3-172b-instruct3	95.48	1.67
llm-jp/llm-jp-3-8x1.8b-instruct2	86.13	7.56
llm-jp/llm-jp-3-8x1.8b-instruct3	92.20	2.20
llm-jp/llm-jp-3-8x13b-instruct2	88.63	6.01
llm-jp/llm-jp-3-8x13b-instruct3	94.35	1.55

Risks and Limitations

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

Send Questions to

llm-jp(at)nii.ac.jp

License

Apache License, Version 2.0

How to cite

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

@inproceedings{
    nakamura2025dropupcycling,
    title={Drop-Upcycling: Training Sparse Mixture of Experts with Partial Re-initialization},
    author={Taishi Nakamura and Takuya Akiba and Kazuki Fujii and Yusuke Oda and Rio Yokota and Jun Suzuki},
    booktitle={The Thirteenth International Conference on Learning Representations},
    year={2025},
    url={https://openreview.net/forum?id=gx1wHnf5Vp}
}

Model Card Authors

The names are listed in alphabetical order.

Hirokazu Kiyomaru, Takashi Kodama and Taishi Nakamura.