🔹 Key Highlights:
- 12% Fewer Parameters: nyun-llama3-62B comprises approximately 12% fewer parameters than the popular Llama-3-70B.
- Intact Performance: Despite having fewer parameters, our model performs at par if not better, and occasionally outperforms, the Llama-3-70B.
- No Fine-Tuning Required: This model undergoes no fine-tuning, showcasing the raw potential of our optimization techniques.
Pipeline and Collaboration
For insights into the pipeline and the list of methods used to optimize these models, check out our PruneGPT repository (https://github.com/nyunAI/PruneGPT). We invite companies and organizations interested in joining forces with us to release more such open-source variants to reach out at contact@nyunai.com.
Model Performance
| Dataset | Nyun-Llama3-62B | Meta-Llama3-70B | Meta-Llama2-70B | MBZUAI K2-65B |
|---|---|---|---|---|
| MMLU (5-shot) | 78.9 | 79.5 | 69.7 | 67.9 |
| Winogrande (5-shot) | 83.3 | 83.1 | 81.8 | 77.0 |
| BoolQ (0-shot) | 85.3 | 79.0 | 73.1 | 83.0 |
| Hellaswag (10-shot) | 85.8 | 88.0 | 86.9 | 85.5 |
| Arc Challenge (25-shot) | 65.9 | 68.8 | 67.2 | 64.8 |
| GSM8K (5-shot) | 70.9 | 76.9 | 52.6 | 50.2 |
| Average | 78.4 | 79.2 | 71.9 | 71.4 |