JonaPoka/Qwen3.5-4B-Finnish

Overview

JonaPoka/Qwen3.5-4B-Finnish is an experimental 4.5 billion parameter instruction fine-tune of the Qwen/Qwen3.5-4B base model, developed by JonaPoka. Unlike continued pre-trains, this is a merged full checkpoint with LoRA baked in, focusing on Finnish language capabilities. It is a text-only model, with its vision tower frozen, and does not include DPO (Direct Preference Optimization).

Key Capabilities

• Finnish Language Generation: Designed to answer in Finnish without defaulting to English.
• Short EN→FI Translation: Capable of performing brief English to Finnish translations.
• Everyday Finnish Drafting: Useful for generating common Finnish texts like cancellation emails or messages.
• Wikipedia-style Explanations: Can produce explanatory content in a style similar to Wikipedia articles.
• Morphology and Register: Training stages focused on improving Finnish register and handling of cases, though some morphological rules (like partitive explanations) remain challenging for the model.

Good for

• Generating Finnish text for general purposes.
• Drafting informal Finnish communications.
• Translating short English phrases or sentences into Finnish.
• Creating explanatory content in Finnish.

Limitations

• Morphology: May still confuse Finnish cases, particularly in rule explanations for the partitive.
• Historical Accuracy: Can invent extra historical details, especially for open-ended prompts.
• Scientific Explanations: Open-ended scientific explanations can be clumsy or inaccurate.
• Capacity: As a 4.5B parameter model, it has less capacity for complex Finnish morphology compared to larger models.
• No DPO: Lacks preference tuning due to technical constraints during development.