Post-training Danish Foundation Models

Vladimir Salnikov · Aarhus University · ongoing

Ongoing research at Aarhus University on post-training Danish foundation models, with a focus on RLAIF (RL from AI feedback) variants adapted to low-resource languages.

The constraint

The bottleneck for Danish post-training isn’t compute — it’s coverage and quality of preference and instruction data. Standard recipes assume an abundant pool of human preference judgements and instruction-tuning conversations. For Danish, both are scarce, and naive English→Danish translation introduces its own distributional artifacts.

What I work on

• RLAIF pipelines that produce useful preference signal at Danish-scale data, including adjustments to the judge prompt and pairwise sampling to reduce English-bias leak-through.
• Synthetic instruction-tuning datasets for Danish, built to fit the same chat-messages schema used by my open-source Scandinavian SFT release. Reusing one schema makes ablations across post-training stages much easier.

This page will track outputs as they appear — preprints, datasets, and write-ups on what worked and what didn’t.