Qwen3-Embedding-Scandi-0.6B

Fine-tuned version of Qwen/Qwen3-Embedding-0.6B for Scandinavian text embeddings (Danish, Norwegian, Swedish).

---

Model Summary

• Base model: Qwen/Qwen3-Embedding-0.6B
• Architecture: Transformer-based embedding model (0.6B parameters)
• Fine-tuning: LoRA + Swift, merged into base weights
• Task: Sentence and document embeddings for retrieval, clustering, and semantic similarity
• Languages: 🇩🇰 Danish, 🇸🇪 Swedish, 🇳🇴 Norwegian

---

Intended Use

This model is intended for representation learning tasks such as:

• Semantic search
• Text clustering
• Document retrieval
• Reranking pipelines

Not recommended for text generation.

---

Training Details

• Dataset: DDSC/nordic-embedding-training-data Scandinavian corpora (mixed Danish, Norwegian, Swedish texts)
• Training framework: Swift with LoRA adapters
• Loss function: InfoNCE

---

Checkpoints

• LoRA weights merged into the base model.
• SafeTensors format used for efficiency.
• Tokenizer from base model copied for compatibility.

---

Limitations & Bias

• Limited to Scandinavian languages (other languages may work poorly).
• Embeddings are sensitive to domain shift (best results on text similar to training data).
• As with all language models, embeddings may encode societal biases present in the training data.

---

Acknowledgements

• Qwen Team for releasing the base model.
• Swift for training utilities.
• Weights & Biases for experiment tracking.
• DDSC for training data.