---
license: odc-by
task_categories:
- text-generation
language:
- ja
tags:
- art
- light-novel
- fiction
- literature
pretty_name: SyosetuNames-3.5M
size_categories:
- 1M<n<10M
configs:
- config_name: default
  data_files:
  - split: clean
    path: syosetu-names.parquet
  - split: raw
    path: syosetu-names-raw.parquet
---

# SyosetuNames-3.5M: Japanese Light Novel Character Names Corpus

## Overview

This dataset extracts fictional character names from the publicly available text of novels on the Japanese light novel platform "Shōsetsuka ni Narō" ([syosetu.com](https://syosetu.com/)), containing **approximately 3.5 million unique original and cleaned names**. This dataset aims to provide resources for culturally sensitive Natural Language Processing (NLP) tasks, such as name generation, Named Entity Recognition (NER) model training, and cross-cultural naming pattern analysis.

## Data Details

The data primarily consists of Japanese fictional character names, encompassing traditional Kanji names and Katakana transliterations of Western names (e.g., "リオン・スミス"). It also includes a small number of fictional names from other languages.

The dataset is stored in Parquet file format and includes both the raw extracted results (`syosetu-names-raw.parquet`) and the cleaned data (`syosetu-names.parquet`).

The original text was indirectly obtained through the third-party preprocessed dataset [botp/RyokoAI_Syosetu711K](https://huggingface.co/datasets/botp/RyokoAI_Syosetu711K).

### Raw Version (syosetu-names-raw.parquet)

This version contains the results after initial extraction and deduplication:

1.  **Rough Extraction**: Initial extraction of character names using the [tsmatz/xlm-roberta-ner-japanese](https://huggingface.co/tsmatz/xlm-roberta-ner-japanese) model.
2.  **Initial Deduplication**: Preliminary deduplication of the identified results to eliminate obvious duplicates.

### Cleaned Version (syosetu-names.parquet)

The cleaned version was constructed through the following process, aiming to improve data quality and usability:

1.  **Rule-based Filtering**: Removal of data from the raw version containing invisible characters and non-alphabetic entries.
2.  **Preliminary Cleaning**: Filtering of invalid characters and context-dependent formats (e.g., dialogue quotation marks) using the Qwen2.5-14B-Instruct model.
3.  **Distillation Optimization**: Transferring the cleaning rules to the smaller and more efficient Qwen2.5-1.5B-Instruct model for efficient processing.
4.  **Second Deduplication**: Further deduplication of all cleaned results to minimize redundancy.
5.  **Rule-based Postprocessing**: Removal of remaining non-alphabetic entries and sorting of the data.

## Use Cases

This dataset is suitable for the following scenarios:

*   **Light Novel Character Name Generation**: Providing a rich vocabulary of names for creative tools or models.
*   **Japanese NER Model Training**: Offering a large amount of annotated data to improve the performance of NER models in recognizing fictional character names.
*   **Cross-cultural Naming Pattern Analysis**: Researching the characteristics and trends of fictional character names across different cultural backgrounds.
*   **Creative Writing Assistance**: Providing inspiration and reference for writers.

**Important Note:** This dataset is not recommended for real identity identification, legal document generation, or other scenarios requiring high accuracy. Do not associate fictional names with real individuals.

## Copyright and License

This dataset is released under the [ODC-By 1.0](https://opendatacommons.org/licenses/by/1-0/) license, allowing free use, modification, and distribution of the cleaned name collection with attribution.

**LEGAL RISK DISCLAIMER:** The copyright of the original text belongs to the "Shōsetsuka ni Narō" platform. When using this dataset, please be sure to comply with the original platform's terms of service. The compliance of the third-party preprocessed dataset has not been fully verified, so careful evaluation is recommended before commercial use. This dataset is more recommended for academic research.

## Citation

```bibtex
@dataset{Sunbread_SyosetuNames,
  title = {SyosetuNames-3.5M: Japanese Light Novel Character Names Corpus},
  author = {Sunbread},
  year = {2025},
  url = {https://huggingface.co/datasets/Sunbread/SyosetuNames-3.5M},
  note = {A corpus of Japanese light novel character names.}
}
```

## Feedback

For any technical issues or suggestions for improvement, please submit them through the [dataset repository's Discussions](https://huggingface.co/datasets/Sunbread/SyosetuNames-3.5M/discussions).