mayflowergmbh/Wiedervereinigung-7b-dpo
Wiedervereinigung-7b-dpo by mayflowergmbh is a 7 billion parameter, DPO-aligned merged language model, primarily based on Mistral architecture and fine-tuned with German datasets. It achieves an average score of 7.11 on the mt-bench-de benchmark, demonstrating strong performance in German language tasks, particularly in writing, roleplay, and humanities. This model is specifically optimized for high-quality German text generation and understanding, making it ideal for applications requiring robust German language capabilities.
Loading preview...
Wiedervereinigung-7b-dpo: A German-Optimized LLM
Wiedervereinigung-7b-dpo is a 7 billion parameter language model developed by Mayflower GmbH, created through a DPO-aligned merge of several prominent German-focused Mistral-based models, including those from DiscoResearch, VAGOsolutions, and malteos. The model was further fine-tuned using a German translation of the SlimOrca DPO dataset, with hermeo-7B used for reject results, enhancing its performance in German language tasks.
Key Capabilities
- High German Language Proficiency: Achieves an average score of 7.11 on the
mt-bench-debenchmark, indicating strong performance in German. - Optimized for German Contexts: Specifically designed and trained to excel in understanding and generating German text.
- Strong in Creative and Conversational Tasks: Scores highly in
writing(8.425),roleplay(8.6), andhumanities(9.35) categories onmt-bench-de. - Multiple Quantized Versions: Available in various quantized formats (GPTQ, AWQ, GGUF, EXL2) for efficient deployment, thanks to LoneStriker.
Good For
- Applications requiring high-quality German text generation, such as content creation, chatbots, and virtual assistants.
- Research and development in German NLP, especially for fine-tuning with German datasets using tools like LLaMA-Factory.
- Use cases where models with strong performance in writing, roleplay, and general humanities in German are critical.