Model Overview

picAIso/TARS-8B is an 8 billion parameter language model developed by picAIso. It was created using the TIES merge method from mergekit, combining the strengths of several pre-trained models based on the Llama-3 architecture.

Key Capabilities

• Instruction Following: Built upon MaziyarPanahi/Llama-3-8B-Instruct-v0.9, it inherits strong instruction-following capabilities.
• Enhanced Performance: Integrates NousResearch/Hermes-2-Pro-Llama-3-8B, suggesting improved general performance and reasoning.
• Text Generation: Incorporates nbeerbower/llama-3-gutenberg-8B, which likely contributes to its ability to generate diverse and coherent text.
• Context Window: Supports an 8192 token context length, suitable for handling moderately long inputs and generating detailed responses.

Merge Details

The model's creation involved merging three distinct models:

• Base Model: MaziyarPanahi/Llama-3-8B-Instruct-v0.9
• Merged Components: NousResearch/Hermes-2-Pro-Llama-3-8B and nbeerbower/llama-3-gutenberg-8B

The TIES merge method was applied with specific density and weight parameters for the contributing models, aiming to combine their respective strengths effectively. The merge process utilized float16 for the dtype and included int8_mask in its parameters.