llm-jp-3-172b-instruct3

This repository provides large language models developed by the Research and Development Center for Large Language Models at the National Institute of Informatics.

The development was partially supported by GENIAC.

Model Variants
llm-jp-3-1.8b
llm-jp-3-1.8b-instruct
llm-jp-3-3.7b
llm-jp-3-3.7b-instruct
llm-jp-3-13b
llm-jp-3-13b-instruct
llm-jp-3-172b-beta1
llm-jp-3-172b-beta1-instruct
llm-jp-3-172b-beta2
llm-jp-3-172b-beta2-instruct2
llm-jp-3-172b
llm-jp-3-172b-instruct3

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-3-172b-instruct3")
model = AutoModelForCausalLM.from_pretrained("llm-jp/llm-jp-3-172b-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::
  • llm-jp-3-1.8b: 2.1T
  • llm-jp-3-3.7b: 2.1T
  • llm-jp-3-13b: 2.1T
  • llm-jp-3-172b-beta1: 0.7T
  • llm-jp-3-172b-beta2: 1.4T
  • llm-jp-3-172b: 2.1T

Params	Layers	Hidden size	Heads	Context length	Embedding parameters	Non-embedding parameters
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
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

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 Japanese instruction dataset.
	answer-carefully-002	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. We used a sampled subset.
	wizardlm8x22b-logical-math-coding-sft_additional-ja	A synthetic instruction dataset. We used a sampled subset.
	magpie-sft-v1.0	A synthetic instruction dataset we created.
English	Daring-Anteater	-
	FLAN	We used a sampled subset.
Japanese & English	Synthetic-JP-EN-Coding-Dataset-567k	A synthetic instruction dataset. We used a sampled subset.

Direct Preference Optimization

We used synthetic preference data to improve both the helpfulness and harmlessness of the LLM. The datasets will be made available soon.

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-3-172b-beta1	0.5174	0.4460	0.2556	0.3700	0.6400	0.6100	0.8265	0.5600	0.5720	0.8505	0.0434
llm-jp-3-172b-beta1-instruct	0.5700	0.4306	0.2292	0.4350	0.8433	0.6200	0.8228	0.6820	0.5873	0.8964	0.1529
llm-jp-3-172b-beta2	0.5422	0.3337	0.2725	0.4700	0.7767	0.6900	0.8283	0.5960	0.6133	0.8380	0.0037
llm-jp-3-172b-beta2-instruct2	0.6022	0.5470	0.2665	0.5100	0.8600	0.7000	0.8392	0.6800	0.6346	0.8770	0.1076
llm-jp-3-172b	0.5431	0.4077	0.2662	0.5150	0.7633	0.6700	0.8227	0.5740	0.5686	0.8289	0.0148
llm-jp-3-172b-instruct3	0.6130	0.5173	0.2711	0.5700	0.8733	0.7300	0.8437	0.7280	0.6012	0.8829	0.1121

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-3-172b-beta1-instruct	5.14	2.90	5.30	8.80	2.15	2.45	6.95	7.45	5.15
llm-jp-3-172b-beta2-instruct2	6.72	4.10	6.90	7.60	4.00	6.35	8.70	7.95	8.15
llm-jp-3-172b-instruct3	7.57	4.85	8.55	9.56	3.75	7.6	8.1	8.95	9.2

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

See the LICENSE file.

Model Card Authors

The names are listed in alphabetical order.

Hirokazu Kiyomaru and Takashi Kodama.