Hodfa71/

saga-da-delta-dpo-r1

Preference pairs used to train the Danish Delta-DPO model in the SAGA paper. 7,414 pairs. Each pair has a prompt (first 25–45% of a Danish Wikipedia sentence), a chosen completion and a rejected one. Only pairs with score gap Δ ≥ 0.25 were kept. Completions scored with SpaCy da_core_news_lg: structural validity (verbal ROOT + nsubj), tree depth, and lexical diversity (MATTR).

	Columns




	

Column Description… See the full description on the dataset page: https://huggingface.co/datasets/Hodfa71/saga-da-delta-dpo-r1.