| Property | Value |
|---|---|
| Parameter Count | 14.8B |
| License | Apache 2.0 |
| Supported Languages | English, German, Italian, French, Portuguese, Dutch, Arabic, Spanish |
| Model Type | DPO Fine-tuned |
| Base Model | SauerkrautLM-v2-14b-SFT |
What is SauerkrautLM-v2-14b-DPO?
SauerkrautLM-v2-14b-DPO is an advanced language model that builds upon its SFT predecessor through Direct Preference Optimization (DPO). This model implements a unique three-phase training approach, combining traditional SFT with DPO techniques to enhance both English performance and German language capabilities.
Implementation Details
The model utilizes a sophisticated training procedure involving three distinct phases: two SFT phases targeting 25% and 20% of layers respectively with 0.6B tokens each, followed by a DPO phase targeting 15% of layers with 80M tokens. The model achieves impressive benchmark scores, including 74.12% on IFEval and 50.93% on BBH (3-Shot).
- Spectrum Fine-Tuning with targeted layer approach
- Enhanced function calling with improved German irrelevance handling
- Balanced dataset composition for maintaining multilingual capabilities
- BF16 tensor type for optimal performance
Core Capabilities
- Multilingual support across 8 languages
- Enhanced English language performance while maintaining German proficiency
- Optimized function calling specifically for German language contexts
- Strong performance on various benchmarks including MMLU-PRO (45.75% accuracy)
Frequently Asked Questions
Q: What makes this model unique?
The model's distinctive three-phase training approach and specific optimization for German-English language handling, combined with its specialized function calling capabilities, sets it apart from traditional language models.
Q: What are the recommended use cases?
The model is particularly well-suited for multilingual applications, especially those requiring strong English-German capabilities, function calling, and general language understanding tasks across multiple domains.