ArliAI/Llama-3.1-8B-ArliAI-RPMax-v1.1: Creative Roleplay Model

This model is part of ArliAI's RPMax series, a collection of models specifically fine-tuned for creative writing and roleplay (RP). Based on the Meta-Llama-3.1-8B architecture, this 8 billion parameter variant features a 32768 token context length.

Key Capabilities & Differentiators

• Highly Creative & Non-Repetitive: Trained on a diverse and deduplicated dataset, ensuring the model avoids repetitive character traits or situations, leading to more dynamic and varied roleplay interactions.
• Versatile Character & Scenario Handling: Designed to adapt appropriately to a wide range of characters and scenarios without 'latching on' to specific personalities.
• Optimized Training: The v1.1 iteration includes a fix to prevent unnecessary training of the lm_head and minor dataset adjustments, enhancing its performance for its intended use.
• Efficient Training: Achieved 1 epoch of training in approximately 1 day on 2x3090Ti, utilizing LORA with 64-rank and 128-alpha, resulting in ~2% trainable weights.

Performance & Usage

While optimized for creative RP, the model's general evaluation on the Open LLM Leaderboard shows an average score of 23.64, with specific metrics like IFEval (0-Shot) at 63.59. Users are encouraged to use the Llama 3 Instruct Format for optimal prompting. The model is available in various quantized formats including FP16, GPTQ_Q4, GPTQ_Q8, and GGUF.