Model Overview

grimjim/llama-3-Nephilim-v1-8B is an 8 billion parameter language model derived from Meta Llama 3, created by grimjim using a SLERP merge of two distinct base models. The primary base model, WhiteRabbitNeo/Llama-3-WhiteRabbitNeo-8B-v2.0, was selected for its high MMLU performance and refusal avoidance. It was merged with mlabonne/NeuralDaredevil-8B-abliterated, a model specifically trained for offensive and defensive cybersecurity, using a very low weight (0.001) to subtly influence the output.

Key Capabilities

• Aggressively Creative Text Generation: The merge is noted for its enhanced creativity and acuity in text generation, making it suitable for diverse content creation.
• Refusal Reduction: When used with the recommended Llama 3 Instruct Direct prompt, it aims to reduce refusals during roleplay without compromising overall safety.
• Cybersecurity Influence: The inclusion of a cybersecurity-trained model suggests potential for nuanced understanding or generation in related contexts, though its primary output is general-purpose creative text.

Usage Considerations

• Licensing: The model is subject to the WhiteRabbitNeo Usage Restrictions Extension to the Llama-3 License and CC-BY-NC-4.0, restricting commercial use without an alternative agreement.
• Potential for Harmful Outputs: Users are advised to exercise care, as the model may generate harmful content.

Performance Metrics

Evaluations on the Open LLM Leaderboard show an average score of 21.58, with specific metrics including IFEval (0-Shot) at 42.77 and MMLU-PRO (5-shot) at 31.06. Detailed results are available here.