Overview

Undi95/MLewd-L2-Chat-13B is a 13 billion parameter model built upon the Llama2-chat architecture, specifically engineered for generating explicit, crude, and kinky content. This model is designed to be uncensored, focusing on outputs that include sexual and lewd themes, particularly for roleplay applications. It emphasizes the importance of the conversation's beginning for maintaining context and consistency.

Key Capabilities

• Uncensored Content Generation: Produces outputs with sexual, crude, and kinky themes without typical content filters.
• Roleplay Optimization: Fine-tuned to excel in roleplaying scenarios, especially those involving adult themes.
• Merged LoRA Architecture: Developed by merging several specialized LoRAs, including Undi95/MLewd-L2-13B-Part3, posicube/Llama2-chat-AYT-13B, zattio770/120-Days-of-LORA-v2-13B, royallab/Pygmalion-2-13b-SuperCOT, and Undi95/MMSoul-13b-lora.
• Alpaca Prompt Template: Utilizes the Alpaca prompt format for instructions and responses.

Performance Benchmarks

While optimized for specific content generation, the model's general reasoning capabilities are reflected in its Open LLM Leaderboard evaluation results:

• Avg.: 51.29
• ARC (25-shot): 62.03
• HellaSwag (10-shot): 84.19
• MMLU (5-shot): 58.75
• TruthfulQA (0-shot): 52.84
• Winogrande (5-shot): 77.43
• GSM8K (5-shot): 11.3
• DROP (3-shot): 12.53

Good for

• Users specifically seeking an LLM for generating explicit, adult-themed, or lewd content.
• Roleplay applications where uncensored and kinky outputs are desired.
• Developers or enthusiasts who require a model with a strong focus on sexual and crude narrative generation.