Nabbers1999/L3.3-70B-PippaMaid-2.0-heretic

Hugging Face
TEXT GENERATIONConcurrency Cost:4Model Size:70BQuant:FP8Ctx Length:8kPublished:May 24, 2026License:llama3.3Architecture:Transformer Warm

Nabbers1999/L3.3-70B-PippaMaid-2.0-heretic is a 70 billion parameter LlamaForCausalLM model, a decensored version of Shifusen/L3.3-70B-PippaMaid-2.0, with an 8192-token context length. This model is specifically fine-tuned using Constitutional AI methodology to significantly improve prose quality in uncensored roleplay (RP) scenarios. It targets and eliminates common AI-generated text artifacts like em-dash abuse, italic overload, and cliche phrases, making it ideal for generating high-quality, human-like narrative content.

Loading preview...

Model Overview

Nabbers1999/L3.3-70B-PippaMaid-2.0-heretic is a 70 billion parameter LlamaForCausalLM, a decensored variant of Shifusen/L3.3-70B-PippaMaid-2.0, built upon the Llama 3.3 70B architecture. It features an 8192-token context length and is designed to enhance prose quality in uncensored roleplay (RP) models, rather than focusing on safety alignment. This model is a merged full-weight version, ready for direct inference or GGUF quantization.

Key Capabilities & Differentiators

  • Prose Quality Improvement: Utilizes Anthropic's Constitutional AI methodology to address common AI-generated text issues in RP, such as em-dash abuse, italic overload, synonym dumping, and cliche phrase recycling.
  • Decensored Output: This "heretic" version is specifically modified to reduce refusals, achieving 3/100 refusals compared to the original model's 100/100.
  • Reproducible Training: The model's creation process is fully reproducible, with detailed parameters and methods available.
  • Constitutional Alignment: Trained against a comprehensive "constitution" defining prose quality principles, including strict formatting bans (e.g., zero em-dashes, no italic emphasis in narration) and anti-slop rules (e.g., 50+ banned AI cliche phrases).
  • Optimized for Narrative Flow: Focuses on concrete physical detail, sentence length variation, dynamic intensity, and narrative continuity across multi-turn interactions.

Performance Metrics

  • Refusals: Significantly reduced to 3/100, down from 100/100 in the original model.
  • Targeted Metric Improvement: Aims for less than 1 "slop phrases" per 1k words (from 5-15), 0 em-dashes per 1k words (from 20-40), and less than 5 italics per 1k words (from 30-60).

When to Use This Model

This model is particularly well-suited for applications requiring high-quality, human-like narrative generation in uncensored roleplay or creative writing contexts where avoiding common AI-generated stylistic flaws is critical. Its focus on prose quality makes it an excellent choice for developers seeking to produce more engaging and natural-sounding text.