Heidrun-Mistral-7B-chat: A Leading Danish LLM

Heidrun-Mistral-7B-chat, developed by Mabeck, is a 7 billion parameter chat model fine-tuned from Heidrun-Mistral-7B-base. It leverages Danish instruction datasets like danish-OpenHermes and skoleGPT to excel in conversational AI for the Danish language. This model was trained efficiently using Unsloth and Huggingface's TRL library.

Key Capabilities

• SOTA Danish Performance: Ranks 1st among Danish open-source LLMs on the ScandEval benchmark, also sharing 1st place across all Scandinavian languages.
• Strong Logic and Reasoning: Demonstrates robust performance in logic and reasoning tasks within the Danish language context.
• ChatML Format: Optimized for the ChatML format, ensuring best performance when used with this specific structure.

Benchmarks

Evaluated using ScandEval, Heidrun-Mistral-7B-chat shows competitive results:

• MMLU-da: 35.66% (ranks 3rd)
• DANSK: 50.80% (ranks 3rd=)
• Hellaswag-da: 29.18% (ranks 4th)

Good for

• Developing Danish-language chatbots and conversational agents.
• Applications requiring high-quality Danish text generation and understanding.
• Research and development in Scandinavian natural language processing, particularly for Danish.