dpo-norwegian-interpreted

Dataset Card for DPO Norwegian Interpreted

Dataset Description

This dataset contains Direct Preference Optimization (DPO) pairs translated and culturally adapted into Norwegian from an English source dataset (e.g. ultrafeedback or safe-rlhf).

Origin and Methodology

• Model Used for Synthesis: kimi-k3 (via local endpoint)
• Translation Strategy: The dataset was generated using a specialized prompt designed to prevent "translationese" and direct literal translation. The kimi-k3 model was instructed to heavily localize the prompt and the chosen response to fit Norwegian culture and idioms. Furthermore, it intentionally generated a rejected response that suffers from poor translation, excessive English loan words, or US-centric hallucinations. This provides a high-quality, targeted signal for DPO/RLHF to teach models to favor authentic, culturally grounded Norwegian over robotic dictionary translations.

Structure

The dataset contains the following fields: - prompt: The culturally adapted Norwegian prompt. - chosen: The high-quality, authentic Norwegian response. - rejected: The poorly translated, uncalibrated response. - target_language: The target language of the row. - dataset_source: The source dataset name. - is_safety_focused: Boolean indicating if this is a safety alignment sample.

License

This dataset inherits the license from its source dataset (MIT or Apache 2.0).