LUNA-SOLARkrautLM-Instruct Overview

LUNA-SOLARkrautLM-Instruct is a 10.7 billion parameter model developed by fblgit, building upon the upstage/SOLAR-10.7B-Instruct-v1.0 base model. It has been aligned using DPO (Direct Preference Optimization) and integrated with UNA technology. A key focus during its development was enhancing its German language proficiency through a unique training approach.

Key Capabilities

• Multilingual Proficiency: Supports both English and German, with a particular emphasis on improved German language skills.
• DPO Alignment: Trained with a custom German SauerkrautLM-DPO dataset, augmented with translated portions of HuggingFaceH4/ultrafeedback_binarized and argilla/distilabel-math-preference-dpo.
• Data Augmentation: Utilizes data augmentation techniques to ensure grammatical and syntactical correctness, and a more natural German wording in its responses.
• Contamination-Free Dataset: Rigorous data contamination tests confirm the SauerkrautLM-DPO dataset is free from unwanted data, with results well below 0.9% across ARC, MMLU, TruthfulQA, and GSM8K.

Good For

• Applications requiring a capable instruction-following model in both English and German.
• Use cases where natural and grammatically correct German phrasing is crucial.
• Developers seeking a DPO-aligned model with a strong focus on German language quality.