SnakModel-7B-Instruct: A Danish-Optimized LLM
SnakModel-7B-Instruct is a 7-billion parameter, instruction-tuned language model developed by the NLPnorth research unit at the IT University of Copenhagen. It is built on the Llama 2 architecture and has undergone extensive continuous pre-training on a diverse collection of 350 million Danish documents (13.6 billion words), followed by fine-tuning on 3.7 million Danish instruction-answer pairs. This specialized training makes it highly proficient in the Danish language.
Key Capabilities
- Danish Language Proficiency: Specifically designed and optimized for understanding and generating Danish text.
- Instruction Following: The
instruct variant is fine-tuned to follow instructions, making it suitable for assistant-like chat applications. - Llama 2 Compatibility: Utilizes the Llama 2 instruction template (
[INST] {instruction} [/INST]). - Quantized Version Available: A 4-bit quantized version optimized for Apple Silicon is provided for local deployment.
Good For
- Danish Language Applications: Ideal for any NLP task requiring strong performance in Danish.
- Chatbots and Virtual Assistants: Excels in assistant-like conversational scenarios in Danish.
- Research and Development: Serves as a robust foundation for further research in Danish Natural Language Processing.
Performance Highlights
SnakModel-7B-Instruct (SnakModel-7B_inst) demonstrates strong performance across various Danish benchmarks, achieving the highest average score (56.63) compared to other Llama 2-based models. Notably, it shows leading results in Named Entity Recognition (NER), Sentiment Analysis (Senti), Summarization (Summ), QA, and Text Classification (TM, CT).