Call-3-Nephilim-v3-8B and flame-3-Nephilim-v3-8B-GGUF These are two innovative models released in Hugging Face. Although these models were never explicitly trained for role-playing, they exhibit remarkable capability in this domain, highlighting the potential of “found art” approaches in ai development.
Creating these models involved merging several pre-trained language models using mergekit, a tool designed to combine the strengths of different models. The llama-3-Nephilim-v3-8B model, with 8.03 billion parameters and using BF16 tensor types, was tested with a temp setting of one and a minimum likelihood (minP) of 0.01. This setting allowed the model to lean toward creative results, which can be tuned as desired. Despite initial issues with format consistency, the model’s performance can be improved through appropriate prompt steering and instruction prompting, ensuring more consistent and varied text generation results.
The llama-3-Nephilim-v3-8B-GGUF variant, which also features 8.03 billion parameters, features multiple quantization options including 4-bit, 5-bit, 6-bit, and 8-bit quantizations. This model was tested with the temperature and minP settings of its counterpart. The inclusion of the GGUF quants in the fusion was intended to maintain creativity while also optimizing the model’s performance for role-playing scenarios.
The research used the Arithmetic Task Fusion method, which allowed combining the strengths of several models. The base model for this fusion was the grimjim/Llama-3-Instruct-8B-SPPO-Iter3-SimPO model, complemented by the tokyotech-llm/Llama-3-Swallow-8B-Instruct-v0.1 model with a lower weight. This combination aimed to improve the chain-of-thought capabilities critical for narrative and role-play coherence.
During testing, it was discovered that none of the components of the merged models had been initially designed for role-playing games. However, through rigorous testing, including role-playing interactions and ad hoc testing, the study identified three models that performed exceptionally well in role-playing scenarios. These included the SPPO (Self-Play Preference Optimization) and SimPO (Simple Preference Optimization with a Reference-Free Reward) models. Despite not having been evaluated on the Open LLM Leaderboard, these models demonstrated strong performance in maintaining narrative coherence and character consistency.
The methodology also highlighted the potential of prompt-guidance in the instructional system. This approach can improve the readability and stylistic appeal of text generation and avoid censorship limitations during role-playing. While some glitches were observed, such as incorrect utterance attribution and spontaneous gender changes, the overall performance of the merged models was impressive.
In conclusion, the release of these models in Hugging Face marks a significant contribution by merging models that were not initially intended for RPGs. The research demonstrated that innovative approaches could yield highly effective results. The llama-3-Nephilim-v3-8B and llama-3-Nephilim-v3-8B-GGUF models are a testament to the potential of ai models to adapt and excel in unforeseen applications.
Sana Hassan, a Consulting Intern at Marktechpost and a dual degree student at IIT Madras, is passionate about applying technology and ai to address real-world challenges. With a keen interest in solving practical problems, she brings a fresh perspective to the intersection of ai and real-life solutions.
