You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

tw-local-knowledge-chat

資料集概述

  • 來源:本地電腦端作業,以「名人、韓流、動漫、政治人物、地標、連鎖商店」六類台灣在地/常見主題為種子(seed_id),組成問答對。
  • 用途:訓練模型回答台灣在地知識、時事人物、流行文化等常見提問,SFT / instruction tuning 用。
  • 主要輸出:datasets.jsonl(2,873 筆)

Splits

Split 檔案 說明
train datasets.jsonl 單一訓練 split,包含完整 2,873 筆資料

資料處理流程

  1. 依六大類別(見下方統計)準備種子清單(seed_id),每個種子生成對應問題(question)。
  2. 使用 BROWSER_AGENT(具網路搜尋能力的 agent)產生回答,讓模型能查證即時/最新資訊後再作答,而非單純憑內部知識生成,藉此降低幻覺與過時資訊的風險。
  3. 回答模型主要為 gemma-4-31B-it,少量問題由 gemini-2.5-flashgemini-2.5-pro 生成(見 gen_model 欄位)。
  4. 整理組裝成 messages 對話格式,輸出為 datasets.jsonl

主要檔案對照

階段 檔案 用途
最終資料集 datasets.jsonl HF 可用訓練資料

欄位說明

欄位 說明
question_id 題目唯一編號(如 question-000001)。
seed_id 種子條目編號,格式為 <category>-<序號>(如 politician-001)。
category 六大分類之一:celebrity / kpop / anime / politician / landmark / chain_store
name 該題目對應的主體名稱(人物/地標/店名等)。
question 使用者提問文字。
answer BROWSER_AGENT 產生的回答內容(與 messages 內 assistant 內容相同)。
gen_model 產生 answer 的模型名稱:gemma-4-31B-it / gemini-2.5-flash / gemini-2.5-pro
cn_words 簡體字/大陸用語偵測輔助欄位(污染掃描用,多數為空字串)。
messages 兩輪對話格式:userquestion)+ assistantanswer)。
text 全文純文字拼接版本。

目前資料統計

  • datasets.jsonl:2,873 筆
  • 類別分佈:
    • celebrity(名人):598
    • kpop(韓國流行文化/藝人):598
    • anime(動漫):455
    • politician(政治人物):429
    • landmark(地標):403
    • chain_store(連鎖商店):390
  • 語言:zh-tw

使用方式(最小示例)

import json

with open("datasets.jsonl", "r", encoding="utf-8") as f:
    sample = json.loads(next(f))

print(sample.keys())

限制與備註

  • 回答內容經 BROWSER_AGENT 網路搜尋輔助生成,仍可能存在生成當下網路資訊本身的錯誤或時效性落差。
  • 少量題目(gemini-2.5-flash / gemini-2.5-pro 生成)與多數題目(gemma-4-31B-it 生成)在語氣/回答結構上可能略有差異。

污染修復情況(2026-08-11 ~ 08-12)

  • 修復方式:污染掃描(詞庫驅動,/data/sdg_data/sdg_pipeline/lexicon/lexicon.yaml)→ scan-triage 逐筆人工判讀「真污染 / 誤報」→ 只修真污染;簡繁/大陸用語為逐詞判讀替換,非 OpenCC 整篇轉換(避免把已正確的繁體改壞)
  • 修復範圍:僅修改 assistant 角色內容(user/tool/system 未動,訓練時本就被 mask)
  • 本次修復筆數:簡體/大陸用語詞庫修正 58 筆、中英夾雜修正 30 筆
  • 完整逐筆 before/after 對照:見 cleanup_0811_before_after.jsonl(詞庫修正)、mixed_lang_0812_tw-local-knowledge-chat.jsonl(中英夾雜修正)

補充修復(2026-08-13)

  • 修復類型:跨語系混雜(韓文/日文音節被漢字誤植取代,如 세계觀세계관
  • 本次修復筆數:3 筆
  • 完整逐筆 before/after 對照:見 wedge_fix_0813.jsonl
  • 本輪為針對性清理(跨語系誤植與注音損壞),非全文重新掃描,不代表已排除其他類型污染。
Downloads last month
29