---
thumbnail: https://github.com/rinnakk/japanese-pretrained-models/blob/master/rinna.png
datasets:
- mc4
- wikipedia
- EleutherAI/pile
- oscar-corpus/colossal-oscar-1.0
- cc100
language:
- ja
- en
tags:
- qwen
inference: false
---

# `rinna/nekomata-14b`

![rinna-icon](./rinna.png)

# Overview
We conduct continual pre-training of [qwen-14b](https://huggingface.co/Qwen/Qwen-14B) on **66B** tokens from a mixture of Japanese and English datasets. The continual pre-training significantly improves the model's performance on Japanese tasks. It also enjoys the following great features provided by the original Qwen model.
* The inclusive Qwen vocabulary (vocab size > 150k) enables the model to processs Japanese texts much more efficiently than the previously released [youri series](https://huggingface.co/collections/rinna/youri-7b-654053610cb8e9d8e6289efc).
* The model supports a maximum sequence length of 8192.

The name `nekomata` comes from the Japanese word [`猫又/ねこまた/Nekomata`](https://ja.wikipedia.org/wiki/%E7%8C%AB%E5%8F%88), which is a kind of Japanese mythical creature ([`妖怪/ようかい/Youkai`](https://ja.wikipedia.org/wiki/%E5%A6%96%E6%80%AA)).


* **Library**

    The model was trained using code based on [EleutherAI/gpt-neox](https://github.com/EleutherAI/gpt-neox).

* **Model architecture**

    A 40-layer, 5120-hidden-size transformer-based language model. Please refer to the [Qwen paper](https://arxiv.org/abs/2309.16609) for architecture details.

* **Continual pre-training**

    The model was initialized with the [qwen-14b](https://huggingface.co/Qwen/Qwen-14B) model and continually trained on around **66B** tokens from a mixture of the following corpora
    - [Japanese CC-100](http://data.statmt.org/cc-100/ja.txt.xz)
    - [Japanese C4](https://huggingface.co/datasets/mc4)
    - [Japanese OSCAR](https://huggingface.co/datasets/oscar-corpus/colossal-oscar-1.0)
    - [The Pile](https://huggingface.co/datasets/EleutherAI/pile)
    - [Wikipedia](https://dumps.wikimedia.org/other/cirrussearch)
    - rinna curated Japanese dataset

* **Training Infrastructure**

    `nekomata-14B` was trained on 16 nodes of Amazon EC2 trn1.32xlarge instance powered by AWS Trainium purpose-built ML accelerator chip. The pre-training job was completed within a timeframe of approximately 7 days.

* **Authors**

    - [Tianyu Zhao](https://huggingface.co/tianyuz)
    - [Akio Kaga](https://huggingface.co/rakaga)
    - [Kei Sawada](https://huggingface.co/keisawada)
    
---

# Benchmarking
Please refer to [rinna's LM benchmark page](https://rinnakk.github.io/research/benchmarks/lm/index.html).

---

# How to use the model

~~~~python
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("rinna/nekomata-14b", trust_remote_code=True)

# Use GPU with bf16
# model = AutoModelForCausalLM.from_pretrained("rinna/nekomata-14b", device_map="auto", trust_remote_code=True, bf16=True)

# Use GPU with fp16
# model = AutoModelForCausalLM.from_pretrained("rinna/nekomata-14b", device_map="auto", trust_remote_code=True, fp16=True)

# Use CPU
# model = AutoModelForCausalLM.from_pretrained("rinna/nekomata-14b", device_map="cpu", trust_remote_code=True)

# Automatically select device and precision
model = AutoModelForCausalLM.from_pretrained("rinna/nekomata-14b", device_map="auto", trust_remote_code=True)

text = "西田幾多郎は、"
token_ids = tokenizer.encode(text, add_special_tokens=False, return_tensors="pt")

with torch.no_grad():
    output_ids = model.generate(
        token_ids.to(model.device),
        max_new_tokens=200,
        min_new_tokens=200,
        do_sample=True,
        temperature=1.0,
        top_p=0.95,
        pad_token_id=tokenizer.pad_token_id,
        bos_token_id=tokenizer.bos_token_id,
        eos_token_id=tokenizer.eos_token_id
    )

output = tokenizer.decode(output_ids.tolist()[0])
print(output)
~~~~

---

# Tokenization
The model uses the original Qwen tokenizer. It augments the [`cl100k` tiktoken tokenizer](https://github.com/openai/tiktoken) and has a vocabulary size of 151,936. The inclusive vocabulary helps the model to reach a better tokenization efficiency, especially for Japanese texts.

We compared the `Qwen` tokenizer (as used in `nekomata`) and the `llama-2` tokenizer (as used in `youri`) on different text collections and found that the Qwen tokenizer achieves a much better byte2token rate (i.e. the average number of tokens produced from 1 byte of text) as following. A lower byte2token rate indicates a better tokenization efficiency.


| Tokenizer | Japanese | English | Multilingual |
| --- | --- | --- | --- |
| Qwen | 0.24 | 0.27 | 0.27 |
| llama-2 | 0.40 | 0.29 | 0.36 |

---

# How to cite
~~~
@misc{RinnaNekomata14b, 
    url={https://huggingface.co/rinna/nekomata-14b}, 
    title={rinna/nekomata-14b}, 
    author={Zhao, Tianyu and Kaga, Akio and Sawada, Kei}
}
~~~
---

# Citations
~~~
@software{gpt-neox-library,
    title = {{GPT-NeoX: Large Scale Autoregressive Language Modeling in PyTorch}},
    author = {Andonian, Alex and Anthony, Quentin and Biderman, Stella and Black, Sid and Gali, Preetham and Gao, Leo and Hallahan, Eric and Levy-Kramer, Josh and Leahy, Connor and Nestler, Lucas and Parker, Kip and Pieler, Michael and Purohit, Shivanshu and Songz, Tri and Phil, Wang and Weinbach, Samuel},
    url = {https://www.github.com/eleutherai/gpt-neox},
    doi = {10.5281/zenodo.5879544},
    month = {8},
    year = {2021},
    version = {0.0.1},
}
~~~
---

# License
[Tongyi Qianwen LICENSE AGREEMENT](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT)