Back to glossary

Term

Llama-3-SauerkrautLM-70b

Llama-3-SauerkrautLM-70b is a German-language DPO fine-tune of Meta Llama 3 70B, developed by VAGOsolutions and Hyperspace.ai.

Llama-3-SauerkrautLM-70b — explained in more detail

Llama-3-SauerkrautLM-70b-Instruct is a 2024 German-language variant of Meta Llama 3 70B Instruct. Trained via DPO (Direct Preference Optimization) over one epoch on roughly 70,000 curated, predominantly German datasets. The model is bilingual (German and English), preserves Llama 3’s underlying capabilities, and translates them into idiomatically clean German.

Benchmarks

  • MT-Bench German: 8.61 average (8.73 first turn / 8.50 second turn) — at release one of the highest scores for open German-language models.
  • German RAG LLM Evaluation: 98.0% overall accuracy.
  • Open LLM Leaderboard: 80.98 average (ARC 74.3 / HellaSwag 87.6 / MMLU 81.1 / GSM8K 91.2 / Winogrande 84.7).

Practical context

Available in EXL2, GGUF, and AWQ quantizations — runnable locally on workstation hardware (Q4 quantization fits in 48 GB VRAM). License: Meta Llama 3 Community License — commercial use with restrictions above 700M monthly active users. Typical use cases: German-language RAG systems, government and business chatbots with data protection requirements.

Entdecke mehr

Themenuebersicht