本資料集是為了 lianghsun/Llama-3.2-Taiwan-3B 與 lianghsun/Llama-3.2-Taiwan-3B-Instruct 設計的「自我認知(self-identity)」訓練資料,協助模型在被問及自身定位、訓練資料時序、能力範圍等問題時,能以一致、明確的繁體中文回答。
Dataset Details
Dataset Description
Llama-3.2-Taiwan-Identity 由若干組「種子提示(seed prompt)」延伸而成。每筆樣本紀錄一個具體的事實陳述(例如:模型是以繁體中文為主、知識截止時間、是否為指令微調版本等),用以在指令微調或 DPO 階段強化模型的自我認知。資料規模刻意保持精簡,目的是作為 identity sub-mix 與其他大型對話語料一起混訓,避免淹沒在通用 instruction 資料中。
Curated by: Huang Liang Hsun… See the full description on the dataset page:
https://huggingface.co/datasets/lianghsun/Llama-3.2-Taiwan-Identity.