Scandinavian Instruction-Tuning Datasets

Live demo ↗

A unified set of instruction-tuning datasets covering the four major Scandinavian written languages — Danish, Swedish, Norwegian Bokmål, and Nynorsk — published on Hugging Face. Combines machine-translated SFT pairs with human–LLM conversations in a single chat-messages schema that any modern fine-tuning stack can consume unchanged.

What’s in it

• A consistent messages schema across four languages, so language switching is a data-loader filter rather than a code change.
• Provenance flags for each example (translated vs. organic human–LLM), so you can run experiments that ablate translation noise.
• Hugging Face Datasets readers, with the splits sized to make small-model SFT runs reproducible on a single GPU.

Preprocessing pipeline

The dataset isn’t just a dump — most of the work was the pipeline that produced it:

• Language identification with nb-nordic-lid to filter out cross-contamination between very-close Scandinavian languages (especially Bokmål vs. Nynorsk vs. Danish, which trip up generic LID models).
• Filtering for length, encoding artifacts, and obvious translation failures.
• Normalization into the same chat messages schema regardless of source, so downstream code doesn’t need per-source adapters.

The goal is to give small Scandinavian-language model efforts a more solid starting point than ad-hoc translated SFT mixes.