---
pretty_name: "Japanese Knowledge Base Synthetic Dataset"
license: "cc-by-nc-4.0"
tags: ["text-generation", "japanese", "dataset"]
size_categories: ["100K<n<1M"]
language: ["ja", "en"]
---
# Japanese Knowledge Base Synthetic Dataset

- **Conversations**: 240,585
- **File size**: ~143 MB
- **Format**: JSON array of chat conversations
- **Content**: synthetic Japanese multi-turn conversations for language tasks

## Overview

This dataset contains high-quality synthetic Japanese question-answering pairs, generated from various rich linguistic and encyclopedic sources. It is primarily designed to improve language models' understanding of Japanese grammar, vocabulary, idioms, terminology, and diverse dialectic nuances.

## Dataset Structure

Each sample is a multi-turn chat conversation between a user and an assistant,
stored in OpenAI messages format. Conversations from the in-repo generation
pipeline have between 2 and 10 user→assistant exchanges (four legacy sources —
Bunpro, Law, Slang, 実用日本語表現辞典 — are single-exchange); assistant turns
are grounded in the source dictionary entry.

### Data Fields

- `messages`: An array of turns, each an object with:
  - `role`: `"user"` or `"assistant"` (turns strictly alternate, user first).
  - `content`: The message text.

### Example

```json
{
  "messages": [
    {"role": "user", "content": "What's the counter for アクセサリー?"},
    {"role": "assistant", "content": "アクセサリーは「点（てん）」で数えます。例えば「一点（いってん）」のように使います。"},
    {"role": "user", "content": "In what situations does that apply?"},
    {"role": "assistant", "content": "商品数や所有数などは「点（てん）」と数えます。"},
    {"role": "user", "content": "Are there alternative counters?"},
    {"role": "assistant", "content": "はい、「個（例：一個（いっこ））」でも数えられます。"}
  ]
}
```

## Source Material

The synthetic conversations were generated from the following domains and
merged into `data.json`. Counts are conversations, not turns (see the merge's
`manifest.json` for the exact figures per run):

| Domain | Source | Conversations |
| --- | --- | --- |
| Grammar (JA-EN) | Bunpro | 913 |
| Antonyms | 対義語辞典オンライン | 12,198 |
| Counters | 数え方辞典オンライン | 1,312 |
| Dialects | 全国方言辞典 | 3,650 |
| Encyclopedia | PixivLight | 30,000 |
| Classical language / Kogo | 旺文社 全訳古語辞典 | 30,000 |
| Law | 有斐閣 法律用語辞典 | 14,711 |
| Medical terms | 南山堂医学大辞典 第20版 | 20,000 |
| Names | JMnedict | 29,673 |
| Onomatopoeia | surasura 擬声語 | 1,414 |
| Slang | ネット用語辞典「ネット王子」 | 5,792 |
| General dictionary | 大辞泉 第二版 | 46,522 |
| General expressions | 実用日本語表現辞典 | 44,400 |
| **Total** | | **240,585** |

The large encyclopedic sources (PixivLight, Kogo, Medical, 大辞泉, JMnedict)
are sampled/capped during generation so they do not swamp the smaller
linguistic sources. 大辞泉 sampling prefers entries with multiple senses,
etymology, or usage notes; JMnedict sampling is stratified by name category
(uncategorized names are dropped).

## Notes

The dataset is distributed as a single JSON array in `data.json`. You can load
it with `datasets.load_dataset` or download it directly from the Hub.

## Citation

```bibtex
@misc{japanese_knowledge_base_synthetic_data,
  author = {Loc Nguyen},
  title = {Japanese Knowledge Base Synthetic Data},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.datasets/bunbohue/Japanese-Knowledge-Base-Synthetic-Data}}
}
```