Hugging Face has introduced two new innovative models named llama-3-Nephilim-v3-8B and llama-3-Nephilim-v3-8B-GGUF. Despite not being explicitly trained for roleplays, these models have demonstrated outstanding proficiency in this area, illuminating the possibilities of “found art” strategies in the domain of artificial intelligence (AI) development.
To create these models, several pre-trained language models were converged. The merger was executed using Mergekit, a tool specifically designed to amalgamate the features of different models. The llama-3-Nephilim-v3-8B model, having 8.03 billion parameters, uses BF16 tensor types. To test this model, a temperature setting of one was applied, along with a minimum probability (minP) of 0.01. This permitted the model to generate creative outputs. However, the model experienced some concerns with consistency, which were addressed by applying instruct prompts and prompt steering. This greatly improved the effectiveness of the model, leading to more consistent and diverse text-generation.
The llama-3-Nephilim-v3-8B-GGUF variant, also with 8.03 billion parameters, provides several quantization options, including 4-bit, 5-bit, 6-bit, and 8-bit. These tests were conducted with the same temperature and minP settings as its counterpart. The purpose of incorporating GGUF quants in the merger was to retain creativity while enhancing the model’s effectiveness in roleplay scenarios.
The research capitalized on the task arithmetic merge technique, allowing for the integration of multiple model strengths. The base model for this merge was the grimjim/Llama-3-Instruct-8B-SPPO-Iter3-SimPO, which was combined with the tokyotech-llm/Llama-3-Swallow-8B-Instruct-v0.1 model at a reduced weight for enhanced fluidity of thought and narrational consistency.
During assessments, it was found that none of the models initially created for roleplay were profoundly useful. However, after extensive testing, including ad hoc trials and roleplay interactions, three models were identified to have performed exceptionally well in roleplay scenarios. SPPO (Self-Play Preference Optimization) and SimPO (Simple Preference Optimization with a Reference-Free Reward) were among these extraordinary models. Though not benchmarked on the Open LLM Leaderboard, the models demonstrated an impressive ability to uphold character consistency and narrative coherence.
The research also revealed the promising potential of prompt steering in the instruction system, which could enhance the readability and artistic appeal of text generation. Though some errors were observed, like utterance misattributions and spontaneous gender flips, the merged models’ overall performance was laudable.
Conclusively, the introduction of these two models on Hugging Face signifies substantial progression through the merging of models that were not originally created for roleplay. Hence, demonstrating that such inventive methods can result in highly efficient output. llama-3-Nephilim-v3-8B and llama-3-Nephilim-v3-8B-GGUF models exemplify the scope of AI models to adapt and thrive in unpredictable applications, pioneering new paths in AI.
