Undi95/MLewd-L2-13B

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:13BQuant:FP8Ctx Length:4kPublished:Sep 4, 2023License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Warm

Undi95/MLewd-L2-13B is a 13 billion parameter language model, based on the Llama 2 architecture, specifically fine-tuned for generating "lewd" or uncensored content. It was developed by Undi95 through a series of merges and LoRA applications, combining models like ReMM, Llama-2-13b-chat-limarp-v2-merged, and LLAMA2-13B-Holodeck-1. This model is optimized for creative text generation in adult-oriented contexts, distinguishing it from general-purpose LLMs.

Loading preview...

Overview

Undi95/MLewd-L2-13B is a 13 billion parameter language model built upon the Llama 2 architecture, specifically engineered by Undi95 for generating "lewd" content. The development involved a complex multi-stage merging process, combining several base models and applying LoRAs to achieve its specialized output. The model's creation focused on iterative refinement, with various versions undergoing testing, including an "ERP Ranking Test" to evaluate its performance in explicit conversational contexts.

Key Capabilities

  • Specialized Content Generation: Designed to produce uncensored and adult-oriented text, diverging from the typical safety alignments of many LLMs.
  • Merged Architecture: Integrates components from Undi95/ReMM, Doctor-Shotgun/llama-2-13b-chat-limarp-v2-merged (Llama Chat Uncensored), and KoboldAI/LLAMA2-13B-Holodeck-1, among others.
  • LoRA Integration: Further refined using LoRAs such as nRuaif/Kimiko-v2-13B and zarakiquemparte/PIPPA-ShareGPT-Subset-QLora-13b to enhance its specific content generation capabilities.

Benchmarks

While optimized for specific content, the model's general language understanding capabilities are reflected in its Open LLM Leaderboard scores:

  • Avg.: 46.84
  • ARC (25-shot): 58.28
  • MMLU (5-shot): 54.67
  • HellaSwag (10-shot): 82.32

Good For

  • Applications requiring uncensored or adult-themed creative writing and conversational responses.
  • Developers seeking a model explicitly designed for "lewd" content generation, where typical LLM safety filters are undesirable.

Popular Sampler Settings

Top 3 parameter combinations used by Featherless users for this model. Click a tab to see each config.

temperature
top_p
top_k
frequency_penalty
presence_penalty
repetition_penalty
min_p