deahmed/Qwen3.5-2B-da-sft

VISIONConcurrent Unit Cost:1Model Size:2.3BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 11, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

deahmed/Qwen3.5-2B-da-sft is a 2.3 billion parameter language model, fine-tuned from Qwen/Qwen3.5-2B, specifically adapted for Danish language tasks. It improves performance on the EuroEval Danish test split by an average of 0.9 points compared to its base model, excelling in sentiment analysis, reading comprehension, and civics knowledge. This model is optimized for Danish text processing while largely retaining its English capabilities and the base model's vision components.

Loading preview...

Model Overview

deahmed/Qwen3.5-2B-da-sft is a specialized 2.3 billion parameter language model, an adaptation of the Qwen/Qwen3.5-2B base model. It has been fine-tuned using LoRA with a Danish instruction dataset of approximately 32,000 examples, including synthetic data, EuroEval-formatted examples, and the skolegpt-instruct dataset, alongside some English data to prevent forgetting. The fine-tuned adapter was merged with the base weights at a strength of 0.3 to balance performance across various tasks.

Key Capabilities

  • Enhanced Danish Performance: Achieves an average improvement of 0.9 points on the EuroEval Danish test split compared to the base Qwen3.5-2B model.
  • Specific Task Improvements: Shows notable gains in:
    • Sentiment analysis (AngryTweets): +2.82 MCC
    • Named Entity Recognition (DANSK): +2.06 F1
    • Reading Comprehension (MultiWikiQA-da): +1.66 F1
    • Civics knowledge (Danish Citizen Tests): +3.90 MCC
  • Retained English Ability: While focused on Danish, the model largely preserves the English capabilities of its base.
  • Composite Model: Inherits the composite text and vision capabilities from the base Qwen3.5-2B, though only the text transformer was fine-tuned and evaluated here.

Intended Use and Limitations

  • Good for: Danish text tasks, particularly those covered by the EuroEval benchmark, such as sentiment analysis, NER, and reading comprehension.
  • Consider Alternatives if: Your use case heavily relies on common-sense reasoning, as the model shows a regression on the HellaSwag-da benchmark (-3.41 MCC). For tasks requiring new factual knowledge beyond what's in the training data, continued pretraining on Danish text might be necessary for larger gains.
  • Vision Capabilities: The base model's vision capabilities are present but were not trained or evaluated in this fine-tuning process.

Finetunes
1 models