LLM-jp-3.1 is a series of large language models developed by the Research and Development Center for Large Language Models at the National Institute of Informatics.

Building upon the LLM-jp-3 series, the LLM-jp-3.1 models incorporate mid-training (instruction pre-training), which significantly enhances their instruction-following capabilities compared to the original LLM-jp-3 models.

This repository provides the llm-jp-3.1-8x13b-instruct4 model. For an overview of the LLM-jp-3.1 models across different parameter sizes, please refer to:

For more details on the training procedures and evaluation results, please refer to this blog post (in Japanese).

Checkpoints format: Hugging Face Transformers

Dense model:

MoE model:

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).

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

In the LLM-jp-3.1 series, we performed continuous pre-training based on Instruction Pre-Training. Instruction Pre-Training enhances a model’s ability to follow instructions by continuing pre-training on a large collection of instruction–response pairs. We prepared approximately 90B tokens of instruction–response data and mixed it with our pre-training datasets, conducting continuous pre-training on a total of 400B tokens. Each model was initialized from existing checkpoints (llm-jp/llm-jp-3-1.8b, llm-jp/llm-jp-3-13b, and llm-jp/llm-jp-3-8x13b) and underwent continuous instruction pre-training. Since the LLM-jp-3 series was originally pre-trained on 2.1T tokens, the total pre-training token count amounts to 2.5T tokens.

Details of this training process will be released in a forthcoming paper. The instruction–response dataset used for this training will also be made publicly available.

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

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

For Direct Preference Optimization (DPO), we adopted rejection sampling. Prompts were sampled from the dataset used in SFT, and multiple responses were generated for each prompt. These responses were then scored (by Qwen/Qwen2.5-32B-Instruct), and DPO was performed by treating high-scoring responses as positive examples and low-scoring responses as negative examples.

We conducted DPO in two stages. In the second stage, we additionally used ac-self-inst, a Japanese preference dataset focused on safety.

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

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

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.

llm-jp(at)nii.ac.jp

Hirokazu Kiyomaru and Takashi Kodama.