Hodfa71/normistral-7b-nb-saga-delta-dpo-pairs

---
language:
- 'no'
- nb
license: apache-2.0
task_categories:
- text-generation
tags:
- SAGA
- preference-learning
- Norwegian
- DPO
- grammar
dataset_info:
features:
- name: prompt
dtype: string
- name: chosen
dtype: string
- name: rejected
dtype: string
- name: chosen_score
dtype: float64
- name: rejected_score
dtype: float64
- name: delta
dtype: float64
splits:
- name: train
num_bytes: 4076789
num_examples: 9133
download_size: 2806606
dataset_size: 4076789
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
---

# NorMistral-7B NB SAGA Δ-DPO Preference Pairs

Preference dataset mined for native Norwegian Bokmål grammar training via SAGA (Syntax-Aligned Grammar Adaptation).

## Details
- **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

## Columns
- `prompt`: 6-word Wikipedia NB prefix
- `chosen`: grammatically better completion (SpaCy PS > rejected)
- `rejected`: grammatically worse completion
- `chosen_score`, `rejected_score`: SpaCy parse scores
- `delta`: score gap (chosen - rejected)