VAGOsolutions/SauerkrautLM-v2-14b-DPO

VAGO solutions SauerkrautLM-v2-14b-DPO

DPO Fine-tuned Model - Enhanced DPO-tuned version with focus on English performance and german function calling irrelevance optimization

Introducing SauerkrautLM-v2-14b-DPO – our advanced DPO-tuned version based on SauerkrautLM-v2-14b-SFT!

• Three-phase training approach combining SFT and DPO
• Enhanced English language performance while maintaining German capabilities
• Optimized function calling with improved german irrelevance handling
• Comes with two new community datasets for custom training (release soon)

Table of Contents

1. Overview of all SauerkrautLM-v2-14b Models
2. Model Details
   • Training procedure
3. Released Datasets
4. Evaluation
5. Disclaimer
6. Contact
7. Collaborations
8. Acknowledgement

All SauerkrautLM-v2-14b

Model	HF	EXL2	GGUF	AWQ
SauerkrautLM-14b-v2-SFT	Link	coming soon	coming soon	coming soon
SauerkrautLM-14b-v2-DPO	Link	coming soon	coming soon	coming soon

Model Details

SauerkrautLM-v2-14b-DPO

• Base Model: SauerkrautLM-v2-14b-SFT
• Language(s): English (primary), German
• License: Apache 2.0
• Contact: VAGO solutions

Training Procedure

This model extends our two-phase SFT model with an additional DPO phase, creating a comprehensive three-phase training approach:

Phase 1 & 2 (SFT):

• Identical to SauerkrautLM-v2-14b-SFT training
• Phase 1: 25% layer targeting with 0.6B tokens
• Phase 2: 20% layer targeting with 0.6B tokens

Phase 3 (DPO):

• Spectrum Fine-Tuning targeting 15% of layers
• Training on 80M tokens
• Focus on English performance optimization
• Integration of German performance preservation
• Enhanced german function calling irrelevance handling

Dataset Composition for DPO:

• Extended previous DPO dataset
• New SauerkrautLM-Fermented-GER-DPO dataset (release soon)
• SauerkrautLM-Fermented-Irrelevance-GER-DPO dataset (release soon)
• Carefully balanced to maintain German language capabilities

Released Datasets

As part of this release, we're making parts of two new datasets available to the community in a few days:

SauerkrautLM-Fermented-GER-DPO:

• 3,300 high-quality German training samples
• Multiple judgment criteria for flexible filtering
• Enables customized training approaches
• Comprehensive metadata for sample selection

SauerkrautLM-Fermented-Irrelevance-GER-DPO:

• 2,000 specialized German training samples
• Focus on function calling irrelevance optimization
• Multiple filtering criteria included
• Designed for community experimentation

Objective and Results

This DPO-enhanced version aims to:

• Optimize English language performance
• Maintain German language capabilities
• Improve german function calling irrelevance handling
• Provide valuable training resources to the community

Evaluation

(same diagrams as in SauerkrautLM-v2-14b-SFT model card)

AGIEVAL

GPT4ALL

TRUTHFULQA

OPENLEADERBOARD 2

MMLU 5-shot

Berkeley Function Calling Leaderboard

Please note that our benchmark results in absolute numbers may differ from the Hugging Face Leaderboard due to variations in benchmark evaluation pipelines. However, the relative differences remain consistent.

Disclaimer

We must inform users that despite our best efforts in data cleansing, the possibility of uncensored content slipping through cannot be entirely ruled out. However, we cannot guarantee consistently appropriate behavior. Therefore, if you encounter any issues or come across inappropriate content, we kindly request that you inform us through the contact information provided. Additionally, it is essential to understand that the licensing of these models does not constitute legal advice. We are not held responsible for the actions of third parties who utilize our models.

Contact

If you are interested in customized LLMs for business applications, please get in contact with us via our website. We are also grateful for your feedback and suggestions.

Collaborations

We are also keenly seeking support and investment for our startup, VAGO solutions where we continuously advance the development of robust language models designed to address a diverse range of purposes and requirements. If the prospect of collaboratively navigating future challenges excites you, we warmly invite you to reach out to us at VAGO solutions

Acknowledgement

Many thanks to Qwen for providing such a valuable base model, and to our community for their continued support and engagement.