DiscoLM German 7b v1: German-Optimized LLM
DiscoLM German 7b v1 is a Mistral-based 7 billion parameter large language model developed by DiscoResearch, building upon the EM German model family. Its core focus is on German-language applications, having undergone supervised fine-tuning (SFT) and DPO reinforcement learning on extensive German and English instruction datasets.
Key Capabilities
- High German Proficiency: Optimized for understanding, generating, and interacting with German language content.
- Multilingual Fluency: Preserves strong fluency in English and excels at translation tasks.
- Instruction Following: Trained on a diverse set of instructions for robust performance in various conversational and task-oriented scenarios.
- ChatML Support: Uses ChatML for prompt formatting, ensuring compatibility with OpenAI endpoints and most inference libraries.
- Optional Retrieval Format: Includes a special retrieval format to enhance steerability and reduce hallucinations in RAG applications.
- Experimental Function Calling: Supports structured outputs and function calling, with ongoing improvements planned.
Good For
- German-centric Applications: Ideal for use cases requiring high-quality German text generation and comprehension.
- Everyday Conversational AI: Designed as a reliable alternative to proprietary models for general German language interaction.
- Translation Tasks: Demonstrates strong performance in translating between German and English.
- RAG Systems: The optional retrieval format can be beneficial for reducing hallucinations in retrieval-augmented generation.
While not primarily aimed at beating benchmarks, preliminary MT Bench results for German indicate strong performance, particularly in reasoning, often comparable to or surpassing GPT-3.5-turbo in specific categories. The model prioritizes perceived language quality for native German speakers over raw benchmark scores.