mlabonne/ChimeraLlama-3-8B

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

ChimeraLlama-3-8B is an 8 billion parameter instruction-tuned language model developed by mlabonne. It is a composite model, created by merging four distinct Llama 3-based models using the dare_ties method via LazyMergekit. This approach combines the strengths of:

• NousResearch/Meta-Llama-3-8B-Instruct
• mlabonne/OrpoLlama-3-8B
• Locutusque/Llama-3-Orca-1.0-8B
• abacusai/Llama-3-Smaug-8B

Key Capabilities & Performance

The primary objective of ChimeraLlama-3-8B is to enhance performance beyond the base Llama 3 8B Instruct model. Evaluations conducted using Nous' benchmark suite demonstrate its effectiveness:

• Outperforms Llama 3 8B Instruct: Achieves an average score of 51.58 on the Nous benchmark, surpassing Meta-Llama-3-8B-Instruct's 51.34.
• Strong in Reasoning and Knowledge: Shows competitive scores across various sub-benchmarks, including AGIEval (39.12), GPT4All (71.81), TruthfulQA (52.4), and Bigbench (42.98).

Good for

• General-purpose conversational AI: Its instruction-tuned nature makes it suitable for a wide range of dialogue-based applications.
• Reasoning and knowledge-intensive tasks: The model's performance on benchmarks like AGIEval and TruthfulQA suggests proficiency in these areas.
• Developers seeking an optimized Llama 3 variant: Offers improved performance over the base Llama 3 8B Instruct without increasing parameter count, making it a compelling alternative for similar use cases.