Model Overview

mobiuslabsgmbh/DeepSeek-R1-ReDistill-Qwen-1.5B-v1.1 is a 1.5 billion parameter language model that has been re-distilled from the original DeepSeek-R1-Distill-Qwen-1.5B. This re-distillation process aims to enhance the model's overall performance and accuracy across a range of benchmarks.

Key Performance Improvements

This version shows notable improvements compared to its predecessor, DeepSeek-R1-Distill-Qwen-1.5B, across several critical benchmarks:

• ARC (25-shot): Improved from 40.96 to 41.55
• HellaSwag (10-shot): Increased from 44 to 45.88
• MMLU (5-shot): Rose from 39.27 to 41.82
• TruthfulQA-MC2: Enhanced from 45.17 to 46.63
• Winogrande (5-shot): Grew from 55.49 to 57.7
• GSM8K (5-shot): Significantly improved from 69.9 to 74.3
• Average: Overall average score increased from 49.13 to 51.31

Further improvements are also observed in more challenging benchmarks:

• MMLU PRO (5-shot): Improved from 16.74 to 19.86
• BBH (3-shot): Increased from 35.12 to 37.23

Use Cases

This model is particularly well-suited for applications requiring:

• Enhanced reasoning capabilities: Demonstrated by improvements in ARC, MMLU, and BBH.
• Accurate question answering: Benefiting from better performance in TruthfulQA and MMLU.
• Mathematical problem-solving: Evidenced by the significant gain in GSM8K scores.

Developers can integrate this model using the provided Hugging Face Transformers library, leveraging its improved performance for various natural language processing tasks.