Dataset Card for belle-multiround
本資料集為以 BELLE 系列指令資料為基礎,整理/轉寫之多輪對話(multi-round) 繁體中文版本,可作為繁中對話模型在多輪互動上的補強資料。
Dataset Details
Dataset Description
BELLE 是早期具規模的中文指令資料集系列。其原始版本以簡體中文為主、且多為單輪指令對話。本資料集做了兩件事:
- 多輪化:以 BELLE 的單輪指令為起點,請 LLM 生成自然延伸的後續輪次(追問、澄清、延伸要求),形成 multi-round 結構。
- 繁中化:將內容轉寫為繁體中文,並調整在地用語。
可用於補強模型在多輪對話中的脈絡保持能力。
- Curated by: Huang Liang Hsun
- Language(s) (NLP): Traditional Chinese
- License: cc-by-nc-sa-4.0
Dataset Sources
- Repository: lianghsun/belle-multiround
- Original BELLE: LianjiaTech/BELLE
- Paper: TBA
Uses
Direct Use
- 補強繁中對話模型在多輪互動上的脈絡保持能力。
- 與其他指令對話資料集(如 tw-instruct-500k)混訓使用。
Out-of-Scope Use
- 不適用作為臺灣特定情境之知識來源;原始 BELLE 並非以臺灣為背景設計。
Dataset Structure
{
"conversations": [
{"from": "human", "value": "寫一篇有關「掌聲響起時」的散文。"},
{"from": "gpt", "value": "..."},
{"from": "human", "value": "可以再幫我把開頭改得更具畫面感嗎?"},
{"from": "gpt", "value": "..."}
]
}
Dataset Creation
Curation Rationale
單輪指令資料常使模型在多輪互動中失去上下文。本資料集以 BELLE 的高覆蓋指令為起點,補上自然的延伸對話,協助模型在多輪情境下穩定回應。
Source Data
Data Collection and Processing
- 取得 BELLE 系列指令資料。
- 以 LLM 為每筆指令生成 1–N 輪自然延伸對話。
- 將內容轉寫為繁體中文,調整在地用語。
Who are the source data producers?
- 原始指令由 BELLE 團隊(LianjiaTech)整理。
- 多輪延伸與繁中化由 LLM 自動完成。
Annotations
Annotation process
無人工標註;多輪延伸由 LLM 生成。
Who are the annotators?
無人工標註者。
Personal and Sensitive Information
資料為合成對話,不含個人識別資料。
Bias, Risks, and Limitations
- 內容主題以原始 BELLE 為基礎,可能仍偏向中國大陸語境的場景。
- 多輪延伸由 LLM 生成,部分樣本可能在第二輪後出現重複或離題。
Recommendations
建議與臺灣本土語境之對話資料集混訓,以平衡語境差異。
Citation
@misc{belle_multiround,
title = {belle-multiround: Multi-round Conversational Variant of BELLE in Traditional Chinese},
author = {Huang, Liang Hsun},
year = {2024},
howpublished = {\url{https://huggingface.co/datasets/lianghsun/belle-multiround}},
note = {Derived from LianjiaTech/BELLE.}
}
Dataset Card Authors
Dataset Card Contact
- Downloads last month
- 21