Model Overview

The mobiuslabsgmbh/DeepSeek-R1-ReDistill-Qwen-7B-v1.1 is a 7.6 billion parameter language model, a re-distilled version of the original DeepSeek-R1-Distill-Qwen-7B. This re-distillation process aims to improve the model's overall performance across a range of benchmarks.

Performance Enhancements

This model demonstrates notable improvements over its predecessor in several key areas:

• MMLU (5-shot): Achieves 59.53%, up from 56.75%.
• TruthfulQA-MC2: Scores 47.7%, an increase from 45.76%.
• Winogrande (5-shot): Reaches 61.8%, compared to 60.38%.
• GSM8K (5-shot): Shows significant improvement with 83.4%, up from 78.85%.
• GPQA (0-shot): Improves to 34.99% from 30.9%.
• MMLU PRO (5-shot): Increases to 31.02% from 28.83%.
• MUSR (0-shot): Jumps to 44.42% from 38.85%.
• BBH (3-shot): Shows a substantial gain to 51.53% from 43.54%.

While most metrics show improvement, ARC (25-shot) and IfEval (0-shot) strict scores are slightly lower than the base model, indicating a trade-off in specific areas. The model maintains a large context window of 131,072 tokens.

Optimized Inference

The model supports integration with HQQ for optimized inference, potentially running approximately 3.5 times faster. This allows for efficient deployment and generation, especially for applications requiring high throughput or reduced latency.