L3 8 B Lunaris v1

Introduction

The L3-8B-Lunaris-v1 is a generalist and roleplaying model based on Llama 3. It is designed to enhance creativity and logic beyond its predecessor, Stheno v3.2. This model is the result of merging multiple models, which has proven more effective than single dataset training in personal experimentation.

Architecture

The model merges several constituent models, each contributing to the overall capabilities in different aspects such as roleplaying and storytelling. The merge configuration specifies various parameters like density and weight for each model to achieve the desired balance.

Mergekit Configuration

• Base Model: meta-llama/Meta-Llama-3-8B-Instruct
• Constituent Models:
  • meta-llama/Meta-Llama-3-8B-Instruct
  • crestf411/L3-8B-sunfall-v0.1
    • Density: 0.4, Weight: 0.25
  • Hastagaras/Jamet-8B-L3-MK1
    • Density: 0.5, Weight: 0.3
  • maldv/badger-iota-llama-3-8b
    • Density: 0.6, Weight: 0.35
  • Sao10K/Stheno-3.2-Beta
    • Density: 0.7, Weight: 0.4

Parameters

• Merge Method: ties
• INT8 Mask: true
• Rescale: true
• Normalize: false
• Data Type: bfloat16

Training

The model's parameters and settings, such as temperature and min_p, are derived from extensive personal experimentation. The merging of models from different datasets is considered more effective than combining them within a single dataset.

Guide: Running Locally

1. Prerequisites:

   • Install Python and necessary libraries.
   • Download the model weights and merge configuration.

2. Installation:

   • Clone the repository containing the model.
   • Install dependencies using a package manager like pip.

3. Execution:

   • Load the model using the provided configuration.
   • Run inference with a suitable input format.

Suggested Cloud GPUs

For optimal performance, consider using cloud-based GPUs such as NVIDIA A100 or V100.

License

The L3-8B-Lunaris-v1 model is released under the llama3 license.