VAGOsolutions/Llama-3-SauerkrautLM-70b-Instruct

VAGO solutions Llama-3-SauerkrautLM-70b-Instruct

Llama-3-SauerkrautLM-70b-Instruct is a 70 billion parameter instruction-tuned model, a collaborative effort between VAGO Solutions and Hyperspace.ai. It is a fine-tuned version of meta-llama/Meta-Llama-3-70B-Instruct, specifically aligned using DPO (Direct Preference Optimization).

Key Capabilities & Features

• Multilingual Proficiency: Enhanced for both German and English, with specific improvements in German language capabilities through curated data.
• DPO Fine-Tuning: Trained for 1 epoch with 70k data points using DPO, contributing to its alignment and performance.
• Performance Benchmarks: Achieves an average score of 80.98 on the Open LLM Leaderboard, including:
  • ARC (25-shot): 74.31
  • MMLU (5-shot): 81.09
  • GSM8K (5-shot): 91.20
• MT-Bench Scores: Demonstrates strong conversational abilities with an average MT-Bench score of 8.688 for English and 8.6125 for German.
• German RAG Evaluation: Shows high accuracy in German RAG tasks, with an overall average of 0.980.

Use Cases

This model is particularly well-suited for applications requiring high-quality language generation and understanding in both English and German, especially in scenarios where robust German language performance is critical, such as customer support, content creation, or information retrieval systems.