import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("llm-jp/llm-jp-3-7.2b")
model = AutoModelForCausalLM.from_pretrained("llm-jp/llm-jp-3-7.2b", device_map="auto", torch_dtype=torch.bfloat16)
text = "自然言語処理とは何か"
tokenized_input = tokenizer.encode(text, add_special_tokens=False, 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

Params	Layers	Hidden size	Heads	Context length	Embedding parameters	Non-embedding parameters
150M	12	512	8	4096	101,874,688	50,344,448
440M	16	1024	8	4096	203,749,376	243,303,424
980M	20	1536	8	4096	305,624,064	684,258,816
1.8b	24	2048	16	4096	407,498,752	1,459,718,144
3.7b	28	3072	24	4096	611,248,128	3,171,068,928
7.2b	32	4096	32	4096	814,997,504	6,476,271,616
13b	40	5120	40	4096	1,018,746,880	12,688,184,320
172b	96	12288	96	4096	2,444,992,512	169,947,181,056

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

Instruction tuning

The models have been fine-tuned on the following datasets.

Language	Dataset	description
Japanese	ichikara-instruction-004-002	A manually constructed instruction dataset
	answer-carefully-002	A manually constructed instruction dataset focusing on LLMs' safety
	ichikara-instruction-format	A small amount of instruction dataset edited from ichikara-instruction, 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_additional-ja	A synthetic instruction dataset.
	Synthetic-JP-EN-Coding-Dataset-567k	A synthetic instruction dataset. We used sampled one.
English	FLAN	We used sampled one.

Evaluation

llm-jp-eval (v1.3.1)

We evaluated the models using 100 examples from the dev split.

Model name	average	EL	FA	HE	MC	MR	MT	NLI	QA	RC
llm-jp-3-1.8b	0.3767	0.3725	0.1948	0.2350	0.2500	0.0900	0.7730	0.3080	0.4629	0.7040
llm-jp-3-1.8b-instruct	0.4596	0.4280	0.1987	0.3250	0.3300	0.4200	0.7900	0.3520	0.4698	0.8224
llm-jp-3-3.7b	0.4231	0.3812	0.2440	0.2200	0.1900	0.3600	0.7947	0.3800	0.4688	0.7694
llm-jp-3-3.7b-instruct	0.5188	0.4191	0.2504	0.3400	0.5000	0.5800	0.8166	0.4500	0.4881	0.8247
llm-jp-3-7.2b	0.5057	0.4062	0.2678	0.3450	0.5800	0.4300	0.8083	0.3480	0.5528	0.8136
llm-jp-3-7.2b-instruct	0.5888	0.4282	0.2659	0.4350	0.8900	0.5800	0.8250	0.4860	0.5565	0.8330
llm-jp-3-13b	0.5802	0.5570	0.2593	0.4600	0.7000	0.6300	0.8292	0.3460	0.5937	0.8469
llm-jp-3-13b-instruct	0.6168	0.5408	0.2757	0.4950	0.9200	0.7100	0.8317	0.4640	0.4642	0.8500

Japanese MT Bench

We evaluated the models using gpt-4-0613. Please see the codes for details.

Model name	average	coding	extraction	humanities	math	reasoning	roleplay	stem	writing
llm-jp-3-1.8b-instruct	4.93	1.50	4.70	7.80	1.55	2.60	7.80	6.10	7.40
llm-jp-3-3.7b-instruct	5.50	1.95	4.05	8.25	2.25	4.00	8.80	7.25	7.45
llm-jp-3-7.2b-instruct	5.70	2.95	5.60	7.95	2.80	3.90	8.40	6.15	7.85
llm-jp-3-13b-instruct	6.47	3.15	7.05	9.15	3.75	5.40	8.30	7.50	7.45

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

Model Card Authors

The names are listed in alphabetical order.

Hirokazu Kiyomaru and Takashi Kodama.