ChimeraLlama-3-8B-v2: A Merged Llama 3 Model

ChimeraLlama-3-8B-v2 is an 8 billion parameter language model developed by mlabonne. It is a product of merging six different Llama 3-based models, including instruction-tuned and DPO-optimized variants, using the LazyMergekit tool. This merging strategy aims to combine the strengths of its constituent models, such as NousResearch/Meta-Llama-3-8B-Instruct, mlabonne/OrpoLlama-3-8B, and cognitivecomputations/dolphin-2.9-llama3-8b, among others.

Key Capabilities & Performance

The model's performance is evaluated on the Open LLM Leaderboard, achieving an average score of 19.99. Specific benchmark results include:

• IFEval (0-Shot): 44.69
• BBH (3-Shot): 28.48
• MMLU-PRO (5-shot): 28.54

These metrics indicate its proficiency across various reasoning and knowledge-based tasks. The model leverages a context length of 8192 tokens, making it suitable for processing moderately long inputs.

Good For

• General-purpose instruction following tasks.
• Applications benefiting from a blend of different Llama 3 fine-tunes.
• Scenarios requiring a capable 8B parameter model with a balanced performance profile.