Tonari-no-usagi's picture
Update README.md
a970929 verified
metadata
language:
  - ja
license: cc-by-nc-4.0
task_categories:
  - text-generation
  - question-answering
tags:
  - function-calling
  - tool-use
  - agent
  - japanese
  - synthetic
size_categories:
  - 1k<n<10k

🇯🇵 Japanese Function Calling Dataset (Synthetic)

Robust & High-Quality Tool-Use Dataset for AI Agents

Buy Full Version License: CC BY-NC 4.0

Overview (概要)

This dataset contains Japanese instruction-response pairs designed to train LLMs for Function Calling (Tool Use) capabilities. Unlike simple translation datasets, this dataset is generated to reflect real-world Japanese usage, including:

  • Casual speech (e.g., Tameguchi)
  • Ambiguous instructions (Subject omission)
  • Irrelevant noise (Realistic conversational fillers)

本データセットは、LLM(大規模言語モデル)に日本語でのFunction Calling(ツール利用/関数呼び出し)能力を学習させるために設計された合成データセットです。 単なる翻訳ではなく、「実用的なAIエージェント」の開発を目的としており、以下のようなリアルな日本語の特徴を含んでいます。

  • 多様な話し言葉: 敬語だけでなく、タメ口など。
  • 曖昧な指示: 主語の省略や、倒置法。
  • ノイズ: 文脈と関係のないフィラー(「えーと」など)や余計な情報。

Why this dataset? (特徴)

Feature Description
High Precision strictly validated via Python scripts. Only valid JSON is included.
スクリプトでJSON構文を厳密にチェック済みです。
Robustness Contains various speech styles to prevent the model from overfitting to "textbook Japanese".
「教科書的な日本語」への過学習を防ぐため、崩した表現を多数収録しています。
Agent-Ready Covering 5+ essential domains: Calendar, Weather, Smart Home, Messaging, and Booking.
カレンダー、天気、家電操作、予約など、エージェント開発に必須のドメインを網羅。

Dataset Structure (データ構造)

The data is provided in JSONL format, compatible with standard chat-completion pipelines. データは一般的なチャット形式(JSONL)で提供されます。

Sample Data

{
  "messages": [
    {
      "role": "user",
      "content": "来週の火曜、大阪に出張なんだけど、天気どう?傘いる?"
    },
    {
      "role": "assistant",
      "tool_calls": [
        {
          "function": {
            "name": "get_weather",
            "arguments": "{\"location\": \"Osaka\", \"date\": \"next Tuesday\"}"
          }
        }
      ]
    }
  ]
}

Defined Tools (収録ツール定義)

The dataset assumes the following tools are available: 以下のツール定義に基づいています。

  • manage_schedule: Add/Check/Delete calendar events.
  • get_weather: Get weather forecast.
  • control_device: Smart home control (Lights, AC, Locks).
  • search_restaurant: Search and book restaurants.
  • send_message: Send texts to contacts.

Access & Licensing (ライセンスと入手方法)

1. Free Preview (Hugging Face)

  • Content: A subset of the dataset (e.g., 500 samples).
  • License: CC-BY-NC 4.0
  • Usage: Research and personal use only. No Commercial Use.
  • 内容: データセットの一部(サンプル)。
  • 用途: 研究、個人開発のみ。商用利用不可。

2. Full Version (Gumroad)

  • Content: Full dataset (10,000+ samples).
  • License: Commercial Use Allowed. You can use this data to train your proprietary models and sell them.
  • Link: Available on Gumroad
  • 内容: フルセット(10,000件以上)。
  • 用途: 商用利用可能。 自社モデルの学習に使用し、そのモデルを販売することができます。

🛠 Generation Process (作成プロセス)

  1. Schema Definition: Defined typical API schemas for AI agents.
  2. Generation: Japanese prompts (Standard, Casual) and corresponding JSON.
  3. Validation: Automatically parsed every JSON output. Invalid JSON or hallucinated parameters were discarded.

Disclaimer (免責事項)

While we strive for high accuracy, this is synthetic data. Please verify important outputs in your specific application context. 高精度を目指して作成されていますが、合成データであるため、実際の運用環境での動作を保証するものではありません。


Ready to build Commercial Agents?

To access the full 13,000+ samples with a Commercial License, please visit the link below: 商用利用可能な 13,000件以上の完全版 は、以下のリンクからご購入いただけます:

👉 Japanese Function Calling Dataset (Full Version)