CoolFace
Datasetpublic

lianghsun/tw-instruct-500k-rephrase-260703

Dataset Card for tw-instruct-500k-rephrase-260703 📚 tw-instruct-500k-rephrase-260703 是一個針對繁體中文(台灣繁體、台灣習慣用語)進行優化的高品質指令微調資料集。 本專案利用 Gemini 2.5 Flash 模型配合 Google 搜尋接地技術(Google Search Grounding),針對原始 lianghsun/tw-instruct-500k 資料集進行回譯、重寫與事實性加強(Factuality Grounding)。 Dataset Details Dataset Description 原始的 lianghsun/tw-instruct-500k 資料集包含了大量的指令微調樣本,但在模型回答的繁體中文語境與事實準確性(Factuality)上仍有提升空間。為了建立一個極致優質、完全符合台灣習慣用語、且具備最新網頁搜尋驗證的高事實性(highly… See the full description on the dataset page: https://huggingface.co/datasets/lianghsun/tw-instruct-500k-rephrase-260703.

sourceHugging Facemitupdated 3mo agoView on Hugging Face
0likes8downloads
Dataset Card

Dataset Card for tw-instruct-500k-rephrase-260703

📚 tw-instruct-500k-rephrase-260703 是一個針對繁體中文(台灣繁體、台灣習慣用語)進行優化的高品質指令微調資料集。 本專案利用 Gemini 2.5 Flash 模型配合 Google 搜尋接地技術(Google Search Grounding),針對原始 lianghsun/tw-instruct-500k 資料集進行回譯、重寫與事實性加強(Factuality Grounding)。

Dataset Details

Dataset Description

原始的 lianghsun/tw-instruct-500k 資料集包含了大量的指令微調樣本,但在模型回答的繁體中文語境與事實準確性(Factuality)上仍有提升空間。為了建立一個極致優質、完全符合台灣習慣用語、且具備最新網頁搜尋驗證的高事實性(highly grounded)資料集,我們開啟了這個重寫計畫。

本資料集採用了以下核心技術與規範:

  1. 1.Google 搜尋接地技術(Google Search Grounding):模型在生成回答時會動態檢索 Google 搜尋引擎,將搜尋結果作為參考依據(grounding sources),大幅降低幻覺(hallucination)並提高回答的實時性與事實正確性。
  2. 2.台灣繁體中文與用語規範:嚴格排除任何簡體字、簡體專用詞(如『我們』、『這』等)以及大陸用語(如『軟件』、『信息』等),一律自動過濾並修復為台灣習慣用語(如『軟體』、『訊息』等)。
  3. 3.完全開源與結構化:資料集保留了每次推論時所使用的 Google 搜尋查詢字詞(search_queries)以及引用的參考來源(sources),便於社群進行可解釋性研究與事實性驗證。

Dataset Structure

每筆資料以 JSONL 格式儲存,包含以下欄位:

  • —row_idx (int): 原始 dataset 的列索引。
  • —prompt (string): 輸入的指令 / 提示詞。
  • —original_output (string): 原始 dataset 中的 GPT 回答。
  • —regenerated_output (string): 由 Gemini 2.5 Flash + Google Search Grounding 生成的新版台灣繁體中文回答。
  • —search_queries (list of strings): 搜尋時所使用的關鍵字。
  • —sources (list of dicts): Grounding 所引用的網頁來源列表(包含標題 title 與網址 uri)。
  • —timestamp (string): 生成時的時間戳記。

範例格式:

json
{
  "row_idx": 499982,
  "prompt": "你桌面上有什麼常見的物品?",
  "original_output": "我作為一個語言模型...",
  "regenerated_output": "我沒有實體的「桌面」...",
  "search_queries": [],
  "sources": [],
  "timestamp": "2026-07-03T04:47:31.944414Z"
}

Dataset Creation

Curation Rationale

為了讓繁體中文模型在指令微調(Instruction Tuning)階段能夠學習到最道地、符合台灣用語習慣的知識,並且在面對實時或事實性問題時能提供高準確度、無幻覺的回答。

Source Data

  • —原始資料集:lianghsun/tw-instruct-500k
  • —生成模型:gemini-2.5-flash
  • —資訊來源:Google Search Grounding

Language & Localization

本資料集採用了專屬的字詞偵測與修正邏輯,確保過濾並替換非台灣用語的內容。

Citation

bibtex
@misc{tw-instruct-500k-rephrase-260703,
  title        = {tw-instruct-500k-rephrase-260703: A High-Factuality Rephrased Dataset with Google Search Grounding},
  author       = {Liang Hsun Huang, (TBA)},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/lianghsun/tw-instruct-500k-rephrase-260703}},
  note         = {Rephrased from tw-instruct-500k using Gemini 2.5 Flash and Google Search Grounding.}
}

Dataset Card Authors

Liang Hsun Huang

Dataset Card Contact

Liang Hsun Huang