Phoenix-7B: German-Optimized DPO Model

Phoenix-7B is a 7 billion parameter language model developed by Matthias Uhlig, specifically fine-tuned for the German language. It is based on the LeoLM/leo-mistral-hessianai-7b architecture and utilizes Direct Preference Optimization (DPO) following the alignment-handbook process. A key differentiator is its training on German translations of HuggingFaceH4/ultrachat_200k and HuggingFaceH4/ultrafeedback_binarized datasets, created using haoranxu/ALMA-13B.

Key Capabilities & Performance

• German Language Proficiency: Optimized for German, addressing the limitations of other models like Mistral in this language.
• DPO-Driven Performance: The DPO training enables Phoenix-7B to compete with and, in some areas, surpass larger models.
• MT-Bench-DE Scores: Beats the LeoLM-Mistral model in most categories and notably outperforms LeoLM/Llama-2-70b-chat in roleplay and reasoning on the German MT-Bench benchmark.
• Research Backing: The model's development and methodology are detailed in the paper "PHOENIX: Open-Source Language Adaption for Direct Preference Optimization" (arXiv:2401.10580).

Use Cases

Phoenix-7B is particularly well-suited for applications requiring high-quality German language generation and understanding, especially in conversational AI, content creation, and tasks demanding strong reasoning and roleplay capabilities within a German context.