Undi95/LewdEngine: A Merged LLaMA2-Based Model

Undi95/LewdEngine is a 13 billion parameter language model developed by Undi95, created by merging the LLaMA2-13B-Holomax and Kimiko-v2-13B models with a lora weight of 0.27. This combination aims to leverage the strengths of both base models to offer a versatile language generation capability.

Key Capabilities & Performance

Evaluated on the Open LLM Leaderboard, Undi95/LewdEngine demonstrates a balanced performance across a range of benchmarks. Its average score is 47.92, indicating general proficiency in language understanding and generation tasks. Specific benchmark results include:

• ARC (25-shot): 60.49
• HellaSwag (10-shot): 83.08
• MMLU (5-shot): 54.84
• TruthfulQA (0-shot): 43.63
• Winogrande (5-shot): 74.9
• GSM8K (5-shot): 12.36
• DROP (3-shot): 6.17

These scores suggest a reasonable capacity for common sense reasoning, factual recall, and language comprehension, while also highlighting areas like mathematical reasoning (GSM8K) and complex reading comprehension (DROP) where performance is lower.

Good For

• General text generation: Suitable for a wide array of tasks where a capable 13B parameter model is needed.
• Exploration of merged models: Developers interested in the performance characteristics of models created through merging different LLaMA2 variants.
• Applications requiring balanced performance: For use cases that benefit from a model with decent scores across multiple linguistic and reasoning benchmarks.