Model Overview

grimjim/llama-3-Nephilim-v3-8B is an 8 billion parameter language model created by grimjim through a merge of pre-trained models using the task arithmetic method. Built upon a Meta Llama 3 base, it integrates tokyotech-llm/Llama-3-Swallow-8B-Instruct-v0.1 to enhance its capabilities.

Key Characteristics

• Creative Text Generation: The model is noted for its creative output, with promptsteering implemented to diversify text generation and address common issues observed in Llama 3 8B models.
• Roleplay Capability: Although not specifically trained for roleplay, the model can be effectively utilized for such applications.
• Merge Details: The merge used grimjim/Llama-3-Instruct-8B-SPPO-Iter3-SimPO-merge as its base, combining it with tokyotech-llm/Llama-3-Swallow-8B-Instruct-v0.1 with a weight of 0.1 for the latter.

Performance Metrics

Evaluations on the Open LLM Leaderboard show an average score of 20.54, with specific results including:

• IFEval (0-Shot): 41.74
• BBH (3-Shot): 28.96
• MMLU-PRO (5-shot): 29.02

Usage Notes

The model was tested with a temperature of 1 and minP of 0.01. Users can adjust the temperature to control creativity. Initial format consistency issues can be mitigated with an Instruct prompt, and specific prompt templates are provided for optimal performance.