Zenz V2.5 Small

🚀 zenz-v2.5-small

Zenz-v2.5-small is a conditional language model based on the GPT-2 architecture, specifically designed for the kana-kanji conversion task and intended for use in the neural kana-kanji conversion system "Zenzai".

✨ Features

• Character-level + byte-level BPE tokenizer.
• High performance in the kana-kanji conversion task.
• Demonstrates high performance in context-aware conversion.

Zenz-v2.5 is available in three model sizes:

• zenz-v2.5-medium: A large-scale model with 310M parameters.
• zenz-v2.5-small: A medium-scale model with 91M parameters.
• zenz-v2.5-xsmall: A small-scale model with 26M parameters.

📚 Documentation

Model Description

This model uses ku-nlp/gpt2-small-japanese-char as the base model, which is provided under the CC-BY-SA 4.0 license. This model is also provided under the CC-BY-SA 4.0 license.

Model Sources

This model is designed to be used in conjunction with Zenzai (AzooKeyKanaKanjiConverter).

• Repository: https://github.com/ensan-hcl/AzooKeyKanaKanjiConverter

Data Sources

This model was built using the zenz-v2.5-dataset.

📄 License

This model is provided under the CC-BY-SA 4.0 license.

Acknowledgements

We would like to express our gratitude to Sakura Internet Inc. for their support in providing computing resources during the construction of this model.

We also utilized the following libraries, tools, and language resources to build this model:

• MeCab (https://taku910.github.io/mecab/)
• ipadic-NEologd (https://github.com/neologd/mecab-ipadic-neologd)
• torch (https://pypi.org/project/torch/)
• transformers (https://pypi.org/project/transformers/)
• datasets (https://pypi.org/project/datasets/)
• jaconv (https://pypi.org/project/jaconv/)
• llama.cpp (https://github.com/ggerganov/llama.cpp)
• llm.c (https://github.com/karpathy/llm.c)