Preference dataset mined for native Norwegian Bokmål grammar training via SAGA (Syntax-Aligned Grammar Adaptation).
Base model: norallm/normistral-7b-warm
Oracle: SpaCy nb_core_news_lg
Pairs: 9,133
δ threshold: 0.25 (quality gap filter)
mean chosen score: 0.862
mean rejected score: -0.994
mean delta: 1.856
prompt: 6-word Wikipedia NB prefix
chosen: grammatically better completion (SpaCy PS >… See the full description on the dataset page:
https://huggingface.co/datasets/Hodfa71/normistral-7b-nb-saga-delta-dpo-pairs.
