SauerkrautLM-v2-14b-SFT: Advanced Fine-Tuning for Enhanced Performance

VAGO solutions introduces SauerkrautLM-v2-14b-SFT, a 14.8 billion parameter instruction-tuned model built upon the Qwen/Qwen2.5-14B architecture. This release marks a significant advancement in fine-tuning methodology, employing a novel two-phase Spectrum Fine-Tuning approach to achieve superior performance.

Key Capabilities and Training:

• Two-Phase Spectrum Fine-Tuning: The model undergoes two distinct training phases, each targeting specific layers and data types. Phase 1 (25% layer targeting) and Phase 2 (20% layer targeting) each utilize 0.6 billion tokens.
• Enhanced Mathematical Reasoning: Training includes carefully curated mathematics data, selected using a proprietary classification model, leading to improved mathematical capabilities.
• Robust Function Calling: Specialized function calling data is integrated across both training phases, boosting the model's proficiency in this area.
• Multilingual Performance: The model is trained on high-quality German and English data from both Sauerkraut-v1 and Sauerkraut-v2, ensuring strong performance in both languages.
• Instruction Following & Common-Sense Reasoning: Significant improvements are noted in the model's ability to follow instructions and apply common-sense reasoning.

Evaluation Highlights:

Evaluations across various benchmarks, including AGIEVAL, GPT4ALL, TRUTHFULQA, OPENLEADERBOARD 2, MMLU 5-shot, and the Berkeley Function Calling Leaderboard, demonstrate the model's advancements. On the Open LLM Leaderboard, it achieves an average score of 35.65, with notable results in IFEval (69.64) and MMLU-PRO (46.73).

Good for:

• Applications requiring strong mathematical problem-solving.
• Scenarios needing reliable function calling capabilities.
• Use cases demanding high-quality German and English language processing.
• Tasks benefiting from improved instruction following and common-sense reasoning.