Overview
eren23/slerp-test-turdus-beagle is a 7 billion parameter language model developed by eren23. It is a product of merging two distinct models, udkai/Turdus and mlabonne/NeuralBeagle14-7B, utilizing the slerp (spherical linear interpolation) merge method. This approach combines the strengths of its constituent models, which are based on the OpenPipe/mistral-ft-optimized-1218 architecture.
Key Capabilities
- General Language Understanding: Excels in a broad range of language tasks.
- Reasoning: Achieves 73.55 on the AI2 Reasoning Challenge (25-Shot) and 70.05 on GSM8k (5-shot).
- Common Sense: Scores 88.85 on HellaSwag (10-Shot) and 83.90 on Winogrande (5-shot).
- Knowledge & Factuality: Demonstrates 64.62 on MMLU (5-Shot) and 69.69 on TruthfulQA (0-shot).
- Merge Configuration: The merge process involved specific layer ranges and parameter weighting for self-attention and MLP layers, indicating a fine-tuned combination strategy.
Performance Highlights
On the Hugging Face Open LLM Leaderboard, eren23/slerp-test-turdus-beagle achieves an average score of 75.11, showcasing its robust performance across diverse benchmarks. Detailed evaluation results are available on the Open LLM Leaderboard.
When to Use This Model
This model is suitable for applications requiring a capable 7B parameter model with balanced performance across reasoning, common sense, and general knowledge tasks. Its merged architecture suggests a blend of capabilities from its base models, making it a versatile choice for various NLP applications.