CalderaAI/13B-BlueMethod
CalderaAI/13B-BlueMethod is a 13 billion parameter language model created by CalderaAI, developed through a complex, tiered merging process of multiple instruction-tuned models. This experimental model combines various LLaMA-based architectures, including Metharme, Nous-Hermes, Vicuna-cocktail, Manticore, HyperMantis, and Alpacino, using a custom script that randomizes layer merging percentages. It is designed for experimental prompting, supporting both Alpaca and Vicuna instruct styles to yield diverse results.
Loading preview...
CalderaAI/13B-BlueMethod Overview
CalderaAI/13B-BlueMethod is a 13 billion parameter language model resulting from an intricate, experimental tiered merging process. Developed by CalderaAI, this model combines several instruction-tuned base models and their composites, utilizing a custom script that randomizes the percentage of each layer merged from one model to the next. This unique merging methodology was a warm-up for a larger project, focusing on exploring novel ways to combine existing model strengths.
Key Composition & Merging Process
The model's creation involved three tiers of merges, building upon established 13B models:
- Tier One Merges:
- 13B-Metharme + 13B-Nous-Hermes = 13B-Methermes
- 13B-Vicuna-cocktail + 13B-Manticore = 13B-Vicortia
- 13B-HyperMantis + 13B-Alpacino = 13B-PsychoMantis
- Tier Two Merges:
- 13B-Methermes + 13B-Vicortia = 13B-Methphistopheles
- 13B-PsychoMantis + 13B-BlueMoonRP = 13B-BlueMantis
- Tier Three Merge:
- 13B-Methphistopheles + 13B-BlueMantis = 13B-BlueMethod
Each constituent model, including those from PygmalionAI, NousResearch, reeducator, openaccess-ai-collective, and Digitous, was carefully selected for its potential contribution to the final ensemble.
Use Cases & Capabilities
This model is particularly suited for experimental prompting. Due to its composite nature, it responds well to both Alpaca and Vicuna instruct prompting styles, offering users the opportunity to explore a wide range of outputs and potentially discover interesting and unexpected results. Its design encourages creative and varied interaction rather than adherence to a single instruction format.