Got Compute, but No Data: Lessons From Post-training a Finnish LLM

TL;DR

This work investigates post-training a Finnish–English LLM by translating instruction and preference datasets into Finnish and evaluating SFT and DPO in a multilingual setting. Using Poro-34B, the authors show that a few hundred Finnish instruction samples can yield competitive Finnish instruction-following, and that multilingual data generally outperforms monolingual English data alone. They find that English-focused preference optimization provides cross-lingual benefits, with the best results emerging when data from both languages are used, though gains over English alone are modest. The study contributes open datasets, recipes, and a Finnish evaluation workflow, highlighting data scarcity as a central challenge and proposing translation-based augmentation as a practical path forward for low-resource languages.

Abstract

As LLMs gain more popularity as chatbots and general assistants, methods have been developed to enable LLMs to follow instructions and align with human preferences. These methods have found success in the field, but their effectiveness has not been demonstrated outside of high-resource languages. In this work, we discuss our experiences in post-training an LLM for instruction-following for English and Finnish. We use a multilingual LLM to translate instruction and preference datasets from English to Finnish. We perform instruction tuning and preference optimization in English and Finnish and evaluate the instruction-following capabilities of the model in both languages. Our results show that with a few hundred Finnish instruction samples we can obtain competitive performance in Finnish instruction-following. We also found that although preference optimization in English offers some cross-lingual benefits, we obtain our best results by using preference data from both languages. We release our model, datasets, and recipes under open licenses at https://huggingface.co/LumiOpen/Poro-34B-chat-OpenAssistant

Figures (1)

• Figure 1: Accuracy by instruction category on English and Finnish IFEval of SFT model trained on the en-fi-100pct data mix.